# e2emfd_towards_endtoend_synchronous_multimodal_fusion_detection__23f49401.pdf

E2E-MFD: Towards End-to-End Synchronous Multimodal Fusion Detection

Jiaqing Zhang1, Mingxiang Cao1, Weiying Xie1 , Jie Lei2, Daixun Li1, Wenbo Huang3, Yunsong Li1, Xue Yang4

1The State Key Laboratory of Integrated Services Networks, Xidian University 2University of Technology Sydney 3Southeast University 4Shanghai AI Laboratory

https://github.com/icey-zhang/E2E-MFD

Multimodal image fusion and object detection are crucial for autonomous driving. While current methods have advanced the fusion of texture details and semantic information, their complex training processes hinder broader applications. Addressing this challenge, we introduce E2E-MFD, a novel end-to-end algorithm for multimodal fusion detection. E2E-MFD streamlines the process, achieving high performance with a single training phase. It employs synchronous joint optimization across components to avoid suboptimal solutions associated to individual tasks. Furthermore, it implements a comprehensive optimization strategy in the gradient matrix for shared parameters, ensuring convergence to an optimal fusion detection configuration. Our extensive testing on multiple public datasets reveals E2E-MFD s superior capabilities, showcasing not only visually appealing image fusion but also impressive detection outcomes, such as a 3.9% and 2.0% m AP50 increase on horizontal object detection dataset M3FD and oriented object detection dataset Drone Vehicle, respectively, compared to state-of-the-art approaches.

1 Introduction

Precise and reliable object parsing is critical in fields such as autonomous driving [1] and remote sensing monitoring [2]. Relying solely on visible sensors can lead to inaccuracies in object recognition in challenging environments, like inclement weather conditions. Visible-infrared image fusion [3; 4; 5; 6] as a typical common multimodal fusion (MF) task addresses these challenges by leveraging complementary information from different modalities, leading to the rapid development of various multimodal image fusion techniques [7; 8; 9; 10; 11]. Techniques like CDDFuse [12] and DIDFuse [13] employ a two-step process where a MF network is trained initially, followed by training an object detection (OD) network with the results from the MF network to assess fusion effectiveness separately. Although deep neural networks have significantly enhanced the ability to learn representations across modalities, resulting in promising multimodal fusion outcomes, the focus has predominantly been on producing visually appealing images. This emphasis often overlooks the improvement of downstream high-level visual tasks, such as enhanced object parsing, which remains a substantial hurdle.

Recent studies have devoted into designing joint learning methods that integrate fusion networks with high-level tasks such as object detection [14] and segmentation [15; 16]. The synergy between MF and OD in Multimodal Fusion Detection (MFD) methods has emerged as a vibrant area of research. This partnership allows MF to produce richer, more informative images, enhancing OD performance, while OD contributes valuable object semantic insights to MF, aiming to accurately locate and identify objects in a scene. Typically, MFD networks adopt a cascaded design where joint

Corresponding author

38th Conference on Neural Information Processing Systems (Neur IPS 2024).

Stage 1 Stage 2

Task-alignment

Stage 1 Stage 1

Multimodal Images Detection Result Fusion Image

OD Object Detection Network

MF Multimodal Fusion Network

Figure 1: Comparison of (d) E2E-MFD with existing MF-OD task paradigms (a) Two-Stage (Separate Cascaded), (b) Two-stage (Joint Cascaded) and (c) Multi-stage (Joint Cascaded).

optimization techniques [17] use the OD network to guide the MF network toward creating images that facilitate easier object detection. Notably, Zhao et. al [18] introduced a joint learning method for multimodal fusion detection, incorporating meta-feature embedding from OD to improve fusion by generating semantic object features through meta-learning simulation. Despite these advancements, as highlighted in Figure 1, significant challenges persist: 1) Current optimization approaches rely on a multi-step, progressively joint method, compromising efficiency; 2) These methods overly focus on leveraging OD information for fusion enhancement, leading to difficulty in parameter balancing and susceptibility to local optima of individual tasks. Therefore, the quest for a unified feature set that simultaneously caters to each task remains formidable.

In this paper, we introduce E2E-MFD, an end-to-end algorithm for multimodal fusion detection, designed to seamlessly blend detailed image fusion and object detection from coarse to fine levels. E2E-MFD facilitates the interaction of intrinsic features from both domains through synchronous joint optimization, allowing for a streamlined, one-stage process. To reconcile fine-grained details with semantic information, we propose the novel concept of an Object-Region-Pixel Phylogenetic Tree (ORPPT) coupled with a coarse-to-fine diffusion processing (CFDP) mechanism. This approach is inspired by the natural process of visual perception, tailored to meet the specific needs of MF and OD. Furthermore, we introduce a Gradient Matrix Task-Alignment (GMTA) technique to fine-tune the optimization of shared components, thereby minimizing the adverse impacts traditionally associated with inherent optimization challenges. This ensures an efficient convergence towards an optimal set of fusion detection weights, enhancing both the accuracy and efficacy of multimodal fusion detection.

Our contributions in this paper are highlighted as follows: (1) We present E2E-MFD, a pioneering approach to efficient synchronous joint learning, innovatively integrating image fusion and object detection into a single-stage, end-to-end framework. This methodology significantly enhances the outcomes of both tasks. (2) We introduce a novel GMTA technique, designed to evaluate and quantify the impacts of the image fusion and object detection tasks. This aids in optimizing the training process s stability and ensures convergence to an optimal configuration of fusion detection weights. (3) Through comprehensive experimentation on image fusion and object detection, we demonstrate the efficacy and robustness of our proposed method.

2 Related Work

2.1 Multimodal Fusion Object Detection

Due to the powerful nonlinear fitting capabilities of deep neural networks, deep learning has made significant progress in low-level vision tasks, particularly in image fusion tasks [19; 20; 7; 21; 22; 23; 24]. Early efforts [9; 25; 13; 26; 27; 28] tended to achieve excellent fusion results by adjusting network structures or loss functions, overlooking the fact that image fusion should aim to improve the performance of downstream application tasks. Fusion images with good quality metrics may be suitable for human visual perception but may not be conducive to practical application tasks [16; 29]. Some research has acknowledged this issue, Yuan et al. [30] utilized various aligned modalities for improved oriented object detection [31; 32; 33; 34] to address the challenge of cross-modal weakly misalignment in aerial visible-infrared images. Liu et al. [17] proposing a joint learning method,

pioneered the exploration of the MF and OD combination methods. Then the optimization of the loss function for segmentation [16] and detection [14] has been validated to be effective in guiding the generation of fused images. They consider the downstream task network as an additional constraint to assist the MF network in generating fusion results with clearer objects. Zhao et al. [18] leveraged semantic information from OD features to aid MF and perform meta-feature embedding to generate meta-features from OD features, which are then used to guide the MF network in learning pixellevel semantic information. Liu et al. [15] proposed multi-interactive feature learning architecture for image fusion and segmentation by enhancing fine-grained mapping of all the vital information between two tasks, so that the modality or semantic features can be fully mutual-interactive. However, OD considers the semantic understanding of objects, while MF and segmentation primarily focus on the pixel-level relationship between image pairs. The optimization coupling of detection and fusion tasks becomes more challenging to investigate these complementary differences, from which both fusion and detection can benefit. It is worth mentioning that these visible-infrared multimodal fusion detection methods are usually designed in a cascaded structure with tedious training steps. Researchers lack the adoption of end-to-end architectures, which would enable one-step network inference to generate credible fused images and detection results through a set of network parameters.

2.2 Multi-task Learning

Multi-task learning (MTL) [35; 36] involves the simultaneous learning of multiple tasks through parameter sharing. Prior approaches involve manually crafting the architecture, wherein the bottom layers of a model are shared across tasks [37; 38]. Some approaches tailor the architecture based on task affinity [39], while others utilize techniques such as Neural Architecture Search [40; 37; 41] or routing networks [42] to autonomously discern sharing patterns and determine the architecture. An alternative method typically combines task-specific objectives into a weighted sum [43; 44; 45]. In addition, most approaches (e.g. [46; 47; 48; 49; 50]) aim to mitigate the effects of conflicting or dominating gradients. The approach of explicit gradient modulation [48; 49; 50; 51], has demonstrated superior performance in resolving conflicts between task gradients by substituting conflicting gradients with modified, non-conflicting gradients. Inspired by the above multimodal learning methods, we introduce a Gradient Matrix Task-Alignment method to align the orthogonal components contained in image fusion and object detection tasks thereby effectively eliminating the inherent optimization barrier that exists between two tasks.

3 The Proposed Method

3.1 Problem Formulation

MF-OD task concentrates on generating an image that benefits emphasizing objects with superior visual perception capability. The goal of OD is to find the location and identify the class of each object in an image which can naturally provide rich semantic information along with object location information. Therefore, the motivation of OD-aware MF is to construct a novel infrared and visible image fusion framework that can benefit from the semantic information and object location information contained in OD. For this purpose, we suppose a pair of visible image x RH W Cx and infrared image y RH W Cy. The optimization model is formulated as: min θt L (t, N (x, y; θt)) , (1)

where t represents the output of the different task network N with the learnable parameters θt. L( ) is a constraint term to optimize the network. Previous approaches solely design image fusion or object detection networks in a cascaded way, which can only achieve outstanding results for one task. To produce visually appealing fused images alongside accurate object detection results, we jointly integrate the two tasks into a unified goal synchronously which can be rewritten as:

θu, θd = arg min ωLu (u, Φ (x, y; θu)) + (1 ω)Ld (d, Ψ (x, y; θd)) + S (θ ) , (2) where θ = θs u = θs d, defined as the shared parameters for MF and OD networks. u and d denote the fused image and detection result, which are produced by the MF network Φ( ) and OD network Ψ( ) with the learnable parameters θu and θd. w is a predefined weighting factor to balance the task training. S( ) is a constrained term to jointly optimize the two tasks. In this paper, we regard the S( ) as a feature learning constrained manner and achieve this goal by designing a Gradient Matrix Task-Alignment training scheme.

PFMM (Branch 0)

RFRM (Branch 1)

RFRM (Branch lth)

RFRM (Branch Lth)

Res Block Res Block

Res Block Res Block

Res Block Res Block

Res Block Res Block

Pixel Object

t ,t = θ z z

Diffusion-based

Detection Head

0z 1 tz Tz tz

( ) 0 t q z z

Node1 Coarse-to-fine Diffusion Process (CFDP)

Node1 Coarse-to-fine Diffusion Process Node2 Object-Region-Pixel Phylogenetic Tree (ORPPT)

Synchronous Joint Optimization

Parameter Space

Gradient Matrix Task-Alignment (GMTA)

Orthogonal and

Multimodal Images θ

( ) ( ) ( ) ( ) ( ) u d u d u d arg 1 , min , , ; ( ) , , ; = + + θ θ u x y θ d x y θ θ

Res Block Res Block

Res Block Res Block

Res Block Res Block

Res Block Res Block

Parameter Space

Gradient Dominance

and Conflict

1 SSIM 2 pixel 3 grad u = + +

( ) ,x y 1o lo L o

Figure 2: An overview of the proposed E2E-MFD framework, which consists of a backbone, nodes, and branches. The backbone is utilized to extract multimodal image features. A fine-grained fusion network (ORPPT) and diffusion-based object detection network (CFDP) are optimized by synchronous joint optimization (GMTA) in an end-to-end manner.

3.2 Architecture

Our proposed E2E-MFD is designed with parallel principle, composited by image fusion and object detection sub-network. Details of the whole architecture are shown in Figure 2. A fusion network (ORPPT) and object detection network (CFDP) can sufficiently realize the granularity-aware detail information and semantic information extraction.

Object-Region-Pixel Phylogenetic Tree. The important fact of that humans pay more attention to different regions from coarse to fine for object detection in object scale and image fusion in pixel scale. Inspired by a phylogenetic tree, to simulate humans to study the interactions of hierarchies under different granularity views, we construct an Object-Region-Pixel Phylogenetic Tree (ORPPT) as Φ( ) to extract different features in multiple region scales. Given an image pair x RW H Cx, y RW H Cy, we firstly extract image features f(x) and f(y) by shared parallel backbone to save memory and computing resources. Then these features are added in channel dimension to obtain the final L multimodal image features o1, ..., ol, ..., o L RW1 H1 C1. The parameters in f( ) are shared with the OD network.

Although f(x) and f(y) can describe the characteristics of visible and infrared modalities, they lack insights into the multi-granularity perspective. Therefore, we utilize the branches including the one pixel feature mining module (PFMM) and L region feature refine module (RFRM) to mine the multiple granularities from coarse to fine. The PFMM B0 is set the same with feature fusion block in the Meta Fusion [18] with the input pair x, y. We set 1, 2, ..., l, ...L to denote each region branch. For branch l, a CNN φl( ) is firstly utilized to extract the granularity-wise feature φl(ol) RW2 H2 C2 at the region level. A set of learnable region prompts Rl = rl,m RC2 Ml m=1 are introduced to define the Ml different regions of granularity-wise feature, where rl,m denotes the mth region prompt at branch l. Then the feature vector is mapped into the region mask Al = al,m RW2 H2 Ml m=1 by conducting the dot product between the feature vector and region prompt followed by batch normalization and a Re LU activation:

Al = Re LU(BN(Rl φl(ol)). (3)

Finally, the vector of the object-level feature is weighted by the region mask and further aggregated to form region representation: bi,j l,m(ol) = ai,j l,mφi,j l (ol), (4)

where bl,m(ol) denotes the mth region representation and (i, j) denotes the spatial location. These region-level representations are further concatenated to form the observation Bl(ol) = [bl,1(ol), bl,2(ol), . . . , bl,Ml(ol)] of branch l. These multi-grained attentions concentrate operation on the spatial location information and the extent of the regions which are similar to the task requirements. The B1, B2, ..., BL are up-sampled to keep the consistent spatial size with the

pixel-level fusion features B0. Then, the region-level fusion features B1, B2, ..., BL are assembled by addition operation followed by a Convolution with 1 1 kernel + Re LU to reduce the channel numbers. Highlighted region-level features are extracted by a Convolution with 1 1 kernel + Sigmoid and then injected into pixel-level features by multiplication and addition. Finally, five Convolutions with 3 3 kernel + Re LU layers are constructed to reconstruct the fusion result u.

Coarse-to-Fine Diffusion Process. Diffusion models, inspired by nonequilibrium thermodynamics, are a class of likelihood-based models. Diffusion Det [52] is the first neural network model to utilize the diffusion model for object detection, introducing a novel paradigm that achieves promising results compared to traditional object detection models. Combining diffusion simulation with the diffusion and recovery process of the object box, the Coarse-to-Fine Diffusion Process (CFDP) introduces it as an efficient detection head to assist fusion networks to focus more on object areas. CFDP model defines a Markovian chain of diffusion forward process by gradually adding noise to a set of bounding boxes. The forward noise process is defined as:

q (zt | z0) = N zt | αtz0, (1 αt) I , (5)

which transforms bounding boxes z0 RN 4 to a latent noisy bounding boxes zt for t {0, 1, . . . , T} by adding noise to z0. αt := Qt s=0 αs = Qt s=0 (1 βs) and βs represents the noise variance schedule. During training stage, a neural network Ψθd (zt, t) is trained to predict z0 from zt by minimizing the training objective with ℓ2 loss:

2 Ψθd (zt, t) z0 2 . (6)

At inference stage, bounding boxes z0 is reconstructed from noise z T with the model Ψθd and updating rule in an iterative way, i.e., z T z T . . . z0. In this work, we aim to solve the object detection task via the diffusion model. A neural network Ψθd (zt, t, x, y) is trained to predict z0 from noisy boxes zt, conditioned on the corresponding image pair x, y.

3.3 Loss Function

The total loss is combined with an image fusion loss function Lf and object detection loss Ld. Lf consists of three types of losses, i.e., structure loss LSSIM, pixel loss Lpixel and gradient loss Lgrad . For one fused image, it should preserve overall structures and maintain a similar intensity distribution from source images. To this end, the structural similarity index (SSIM) is introduced in function:

LSSIM = (1 SSIM(u, x)) /2 + (1 SSIM(u, y)) /2, (7)

where LSSIM denotes structure similarity loss. In the fused image, we expect the object regions to have a more significant contrast compared to the background region. Therefore, the object regions need to preserve the maximum pixel intensity and the background region needs to be slightly below the maximum pixel intensity to bring out the contrast between the object and background. The ground-truth bounding boxes of the objects in images are denoted as (xc, yc, w, h) for horizontal boxes and (xc, yc, w, h, θ) for rotated boxes, where (xc, yc) is the center location, w and h are the width and height, θ = angle pi/180, respectively. Based on these ground-truth bounding boxes, we construct the object mask Im, and the background mask is denoted as 1 Im. The object regions pixel loss Lo pixel and the background region pixel loss Lb pixel are formulated as:

Lo pixel = Im (u max(x, y)) 1 , Lb pixel = (1 Im) (u mean(x, y)) 1 , (8)

where 1 stands for the l1-norm. The operator denotes the elementwise multiplication, max( ) denotes the element-wise maximization, and mean( ) denotes the element-wise average operation. Therefore, the object-aware pixel loss Lpixel is defined as:

Lpixel = Lo pixel + Lb pixel . (9)

Besides, gradient information of images always characterizes texture details, thus, we use Lgrad to constrain these textual factors to a multi-scale manner:

ku max kx, ky 2

where denotes gradient operators that calculate by = u G(u) with combination of different Gauss (G) kernel size k. Totally, we obtain Lu = η1LSSIM + η2Lpixel + η3Lgrad .

Figure 3: Visual results of image fusion on M3FD.

3.4 Gradient Matrix Task-Alignment

The MF and OD tasks have distinct optimization objectives. MF primarily emphasizes capturing the pixel-level relationship between image pairs, while OD incorporates object semantics within the broader context of diverse scenes. An inherent optimization barrier exists between these two tasks. We observe that the prevailing challenges in multi-task learning are arguably task dominance and conflicting gradients. We introduce a Gradient Matrix Task-Alignment (GMTA) by presenting the condition number to mitigate the undesired effects of the optimization barrier in task-shared parameters θ which are supposed to be balanced between the MF and OD tasks. The individual task gradients of MF and OD task are calculated by gu = θ Lu and gd = θ Ld in the training optimization process. The gradient matrix can be defined as G = {gu, gd}. In multi-task optimization, a cumulative gradient g = Gw is a linear combination of task gradients and the stability of a linear system is measured by the condition number of its matrix in numerical analysis. Hence the stability of the gradient matrix is equal to the ratio of the maximum and minimum singular values (non-negative) of the corresponding matrix: κ(G) = σmax

σmin . Learning from the Aligned-MTL [51], the condition number is optimal (κ(G) = 1) if and only if the gradients are orthogonal and equal in magnitude which means that the system of gradients has no dominance or conflicts:

κ(G) = 1 < gu, gd >= 1. (11)

The final linear system of gradients defined by ˆG satisfies the optimal condition in terms of a condition number. Thereby, we consider the feature learning constraint S(θ ) can be defined as the following optimization to eliminate instability in the training process:

min ˆG G ˆG 2 F s.t. κ( ˆG) = 1 min ˆG G ˆG 2 F s.t. ˆG ˆG = I. (12)

The problem can be treated as a Procrustes problem and can be solved by performing a singular value decomposition (SVD) to G (G = UΣV T ) and rescaling singular values corresponding to principal components so that they are equal to the smallest singular value:

ˆG = σUV = σGV Σ 1V T , (13)

where, (V , λ) = eigh(G G), (14)

Σ 1 = diag( p

1/λmin), (15)

eigh represents a function for finding eigenvectors V and eigenvalues λ and diag stands for diagonal matrix. λmax and λmin are maximum eigenvalues and minimum eigenvalues from λ.The stability criterion, a condition number, defines a linear system to an arbitrary position scale. To alleviate this ambiguity, we choose the largest scale that guarantees convergence to the optimum: this is a minimal singular value of an initial gradient matrix:

σ = σmin(G) = p

Table 1: Quantitative results of different fusion methods on TNO, Road Scene, and M3FD datasets. The model training (Tr.) and test (Te.) time is counted on an NVIDIA Ge Force RTX 3090. The best result is highlighted.

Task Method M3FD TNO Road Scene Tr. Time Te. Time EN MI VIF EN MI VIF EN MI VIF

DIDFuse[13] 6.13 14.65 1.51 6.30 15.30 1.47 6.67 16.65 1.55 3h9m38s 0.096s U2Fusion[26] 5.66 14.22 1.50 5.78 14.89 1.49 6.25 16.30 1.57 4h8m36s 2.091s PIAFusion[58] 5.75 13.92 1.59 5.05 13.61 1.36 6.37 16.22 1.58 5h35m20s 0.003s Swin Fusion[59] 5.80 13.83 1.58 6.09 14.28 1.55 6.30 15.93 1.60 3h38m5s 0.044s CDDFuse[12] 5.77 13.82 1.58 6.21 15.03 1.49 6.54 16.54 1.57 5h59m59s 0.096s

Tardal[17] 5.72 14.68 1.47 5.87 14.99 1.43 6.72 16.98 1.54 5h36m28s 0.093s Metafusion[18] 6.20 15.19 1.54 6.29 16.03 1.44 6.35 16.76 1.57 6h47m38s 0.002s E2E-MFD 6.36 15.47 1.65 6.40 16.28 1.60 6.79 17.11 1.69 2h50m32s 0.014s

4 Experiments and Analysis

4.1 Dataset and Implementation Details

We conduct experiments on four widely-used visible-infrared image datasets: TNO [53], Road Scene [26], M3FD [17] and Drone Vehicle [3]. TNO and Road Scene are just used to evaluate MF performance. M3FD is adopted to evaluate both MF and OD performance. Road Scene with 37 image pairs, TNO with 42 image pairs and M3FD with 300 pairs are only used for the MF task in the testing stage, and the MF network is trained by the M3FD dataset which is divided into a training set (2,940 image pairs) and a testing set (1,260 image pairs). Besides, Drone Vehicle consists of 28,439 image pairs is utilized to train and test MF and OD for oriented objects. We conduct all the experiments with one Ge Force RTX 3090 GPU, and the code of M3FD is based on Detectron2 [54], while the code of Drone Vehicle is based on MMDetection 2.26.0 [55] and MMRotate 0.3.4 [56]. On the M3FD dataset, the pretrained Diffusion Det is used for the initialization of the OD network. In the training phase, E2E-MFD is optimized by Adam W with a batch size of 1. We set the learning rate to 2.5e 5 and the weight decay as 1e 4. The default training iteration is only 15,000. On the Drone Vehicle dataset, the pretrained LSKNet [57] is used for the initialization of the object detection network, and we fine-tune it for 12 epochs with a batch size of 4. The E2E-MFD is optimized by Adam W and the learning rate and the weight decay is set to 1e 4 and 0.05.

4.2 Main Results

Results on Multimodal Image Fusion. Qualitative results of different fusion methods are depicted in Figure 3. All the fusion methods can fuse the main features of the infrared and visible images to some extent and we can observe two remarkable advantages of our method. First, the significant characteristics of infrared images can be effectively highlighted by our method. Our M3FD fusion image captures the person riding a motorcycle. In comparison with other methods, our method demonstrates high contrast and recognition of the objects. Second, our method preserves rich details from the visible images, including color and texture. Our advantages are evident in the fusion images across the M3FD dataset, such as the clear outline of the white car s rear and a man on a motorcycle. While retaining a substantial amount of detail, our method maintains a high resolution without introducing blurriness. In contrast, other methods fail to achieve these two advantages simultaneously. Sequentially, we provide quantitative results of different fusion methods in Table 1. Our E2E-MFD generally achieves the best metric values. Specifically, the largest average value on MI proves that our method transfers more considerable information from both source images. Values of EN reveal that our results contain edged details and the highest contrast between objects and the background. The large VIF shows our fusion results have high-quality visual effects and small distortion compared with the source images. Moreover, our method achieves the fastest training time to finish the joint learning at one stage which means that faster iterative updates can be done on new datasets. The test time to generate a fused image ranked third.

Results on Multimodal Object Detection. To more effectively evaluate the fusion images and observe their impact on downstream detection tasks, we conduct tests using the baseline detector YOLOv5s on all SOTA methods on the M3FD dataset. We follow the same parameter settings, and the visualization results are shown in Figure 4. The detection results are poor when using only single-modal image inputs, with instances of missed detection, such as the motorcycle and rider next to the car and people on the far right in the image. Almost all fusion methods reduce

Table 2: Quantitative results of object detection on M3FD dataset among all the image fusion methods + detector (i.e. YOLOv5s [60]). means using the fusion images generated by E2E-MFD for object detection training. The best result is highlighted.

Task Method People Car Bus Motorcycle Lamp Truck m AP50 m AP50:95 V/I Infrared 49.3 67.1 72.9 35.8 43.6 61.6 85.3 55.1 Visible 38.1 69.4 75.5 44.4 44.8 63.2 86.3 55.9

DIDFusion[13] 45.8 68.8 73.6 42.2 43.7 61.5 86.2 56.2 U2Fusion[26] 47.7 70.1 73.2 43.2 44.6 63.9 87.1 57.1 PIAFusion[58] 46.5 69.6 75.1 45.4 44.8 61.7 87.3 57.2 Swin Fusion[59] 44.5 68.5 73.3 42.2 44.4 63.5 85.8 56.1 CDDFuse[12] 46.1 69.7 74.2 42.2 44.2 62.7 87.0 56.5

E2E-OD CFT[61] 52.0 68.2 79.2 49.9 45.2 69.6 89.8 60.7 ICAFusion[62] 48.8 68.5 72.3 45.5 43.6 64.7 87.4 57.2

Tardal[17] 49.8 65.4 69.5 46.6 43.7 61.1 86.0 56.0 Meta Fusion[18] 48.4 66.7 70.5 49.1 46.4 59.9 86.7 56.8 Ours (YOLOv5s) 51.0 67.9 69.4 50.2 48.7 61.6 87.9 58.1 Ours (Diffusion Det) 58.5 67.7 79.9 50.3 46.2 70.2 90.3 62.1 Ours (E2E-MFD) 60.1 69.5 81.4 52.2 47.6 72.2 91.8 63.8

Figure 4: Visual results of object detection on M3FD.

missed detection and improve confidence by fusing information from both modalities. Through the design of an end-to-end fusion detection synchronous optimization strategy, we obtain fusion images that are visually and detection-friendly, especially for occluded and overlapping objects, as seen in the blue ellipse with the motorcycle and the overlapping people on the far right in the image. To further assess the quality of fusion images, we conduct a fair comparison between our method and the SOTA methods on YOLOv5s. As shown in Table 2, the MF methods demonstrate a performance improvement compared to single-modal detection, indicating that well-fused images can effectively assist downstream tasks. In contrast, the fusion image we generate has achieved the best performance on YOLOv5s. Additionally, the detection performance of fusion images on Diffusion Det is also impressive, albeit slightly lower than when optimizing fusion and detection tasks

Table 3: Quantitative results of object detection on Drone Vehicle test sets among all the SOTA methods, where means using the fusion images generated by E2E-MFD for object detection training and testing. The best result is highlighted.

Modality Detectors Car Truck Freight Car Bus Van m AP50

Retina Net-OBB [63] 67.5 28.2 13.7 62.1 19.3 38.1 Faster R-CNN-OBB [64] 67.9 38.6 26.3 67.0 23.2 44.6 Gliding Vertex [65] 75.8 46.1 33.8 68.1 38.7 52.5 YOLOv5s-OBB [60] 89.0 53.6 41.9 84.8 32.6 60.4 LSKNet-OBB [57] 89.5 70.0 51.8 89.4 56.9 71.5

Retina Net-OBB [63] 79.9 32.8 28.1 67.3 16.4 44.9 Faster R-CNN-OBB [64] 88.6 42.5 35.2 77.9 28.5 54.6 Gliding Vertex [65] 89.2 59.7 43.0 78.8 43.9 62.9 YOLOv5s-OBB [60] 95.6 57.2 47.5 89.4 35.2 65.0 LSKNet-OBB [57] 90.3 73.3 57.8 89.2 53.2 72.8

UA-CMDet [3] 87.5 60.7 46.8 87.1 38.0 64.0 TSFADet [30] 89.2 72.0 54.2 88.1 48.8 70.4 CALNet [66] 90.3 76.2 63.0 89.1 58.5 75.4 Ours (YOLOv5s-OBB) 96.7 69.9 49.9 92.6 44.5 70.7 Ours (LSKNet-OBB) 90.3 77.0 63.5 89.5 59.0 75.9 Ours (E2E-MFD) 90.3 79.3 64.6 89.8 63.1 77.4

simultaneously with E2E-MFD. Thanks to the collaborative optimization of both tasks, the detection performance is further enhanced. Furthermore, even when compared to end-to-end object detection methods (E2E-OD), our approach demonstrates significant performance improvements. This better underscores the advantages of our training paradigm and the effectiveness of our method.

Results on Multimodal Oriented Object Detection. As shown in Table 3, our fusion detection synchronous optimization strategy achieves the highest accuracy. Furthermore, the outstanding detection performance on YOLOv5s-OBB [60] and LSKNet using the generated fusion images (with at least 5.7% and 3.1% higher AP values compared to single modalities) demonstrate the robustness of our method. This validates the superior quality of the fusion images, indicating that they are not only visually appealing but also provide rich information for the detection task.

4.3 Ablation Studies

Analysis of Gradient Matrix. As described in Section 3.4, the MF and OD tasks pursue different optimization goals. To visualize the task dominance and conflicting gradients, we plot the gradient matrix in the training stage illustrated by Figure 5. We perform a GMTA operation every 1,000 iteration loss updates. Blue represents the gradients of shared parameters computed by the OD loss function, while yellow represents the gradients of shared parameters computed by the MF loss function. During the training process, it can be observed that the gradient values of the OD task are larger and dominant, while that of the MF task are smaller. This may affect the learning process of the fusion task during training. Conversely, the utilization of GMTA effectively mitigates this gradient dominance and conflict, facilitating a balance of shared parameters between MF and OD.

Table 4: The validation of GMTA on M3FD.

Task EN MI VIF m AP50 m AP50:95 MF 6.09 14.90 1.48 / / OD / / / 90.28 62.75 E2E-MFD (w/o GMTA) 6.12 14.70 1.39 90.05 62.60 E2E-MFD (w GMTA) 6.36 15.47 1.65 91.80 63.83

Effect of Gradient Matrix Task Alignment. To verify the effectiveness of GMTA, we compare separate optimizations for MF and OD, as well as joint optimization, w/o and w indicating whether to use GMTA. Specifically, MF represents using only Lu to optimize the fusion network, OD represents using only Ld to optimize the object detection network, and E2E-MFD represents simultaneous optimization of the fusion and detection networks using both Lu and Ld loss functions. The results, as shown in Table 4, indicate that methods incorporating GMTA optimization constraints with shared weights achieve the best results for both MF and OD. This is because MF primarily emphasizes capturing the pixel-level relationship between image pairs, while OD incorporates object semantics within the broader context of diverse scenes. Therefore, optimizing the entire network with shared loss functions may be influenced by local optimal solutions of individual tasks. The accuracy of E2E-MFD (w/o GMTA) shows a slight decrease compared to separately training the detection network. In contrast, GMTA orthogonalizes the gradients of shared parameters corresponding to the two tasks, allowing the joint network to converge to an optimal point with fusion detection weights.

Table 5: Ablation of different MTL methods on M3FD.

Method EN MI VIF m AP50 m AP50:95 E2E-MFD (w/o GMTA) 6.12 14.70 1.39 90.05 62.60 PCGrad[50] 6.13 15.01 1.48 90.59 62.71 CAGrad[47] 6.17 15.05 1.48 90.71 62.74 Nash-MTL[49] 6.29 15.28 1.51 90.91 62.97 E2E-MFD (w GMTA) 6.36 15.47 1.65 91.80 63.83

To compare the effectiveness of different Multi-task learning methods on our algorithm, we selected three robust MTL optimization methods. As shown in Table 5, all MTL methods addressed the conflict between the MF and OD tasks to varying degrees. By introducing the concept of GMTA, we achieved a better balance in the gradient optimization process between the two tasks, resulting in the best performance.

Table 6: Ablation studies of the iteration parameter n on M3FD dataset.

n EN MI VIF m AP50 m AP50:95 500 5.93 14.78 1.58 90.93 62.73 1000 6.36 15.47 1.65 91.80 63.83 1500 6.24 15.08 1.62 91.10 62.96 2000 6.13 14.69 1.45 90.35 62.75

The GMTA process operates during the computation and updating stages of two gradients. The GMTA is performed approximately every n iteration (gradient update), focusing on balancing the independence and coherence of various tasks. Table 6 presents the ablation analysis of the n parameter. Decreasing n initially disrupts task optimization due to frequent alignment, while increasing n becomes crucial when

(a) 1000iteration

(b) 13000iteration

w/o GMTA w GMTA w/o GMTA w GMTA

Figure 5: Visualization of task dominance and conflicting gradients in joint learning of OD and MF.

the network determines task optimization directions. However, excessively large n leads to significant deviations in task paths, making alignment more challenging and negatively impacting performance.

Table 7: Ablation study of the number of ORPPT branches on M3FD.

Branch EN MI VIF m AP50 m AP50:95 0 6.10 15.19 1.54 91.51 63.20 0,1 6.10 15.29 1.54 91.65 62.96 0,1,2 6.19 15.28 1.58 91.73 63.46 0,1,2,3 6.36 15.47 1.65 91.80 63.83 0,1,2,3,4 5.99 14.31 1.40 91.73 63.55

Study of Branches in the Object-Region-Pixel Phylogenetic Tree. We investigate the combination of the pixel feature mining module (0) and the region feature refinement module (1,2,3,4). The results are shown in Table 7. It can be observed that with the increase in the number of branches, the fusion network achieves higher image fusion quality and object detection performance. Region features provide the fusion network with multi-level semantic features of objects. However, when higher-level semantic object information is added, the performance of the fusion network declines. This is because the detailed information contained in the deeper layers of the backbone structure shared with the detection network decreases abruptly, which may affect the fusion network s absorption of these features, thereby influencing pixel-level fusion tasks.

We have implemented the Object-Region-Pixel Phylogenetic Tree (ORPPT) to explore the hierarchical interactions under different granularity views and extract various features across multiple region scales, as introduced in Section 3.2. As shown in Figure 6, the detailed information will decrease as the backbone structure shared by the detection network deepens, resulting in a decrease in detailed information. This may affect the absorption of these features by the fusion network of the shared backbone network, thereby affecting pixel-level fusion tasks. This provides evidence for our analysis of the reasons for Section 4.3 ablation experiment for Object-Region-Pixel Phylogenetic Tree. This illustrates the importance of balancing the semantic information provided between the OD and MF network with pixel-level information.

0 B 1 B 2 B 3 B 4 B

Figure 6: Feature map visualization of various branches in the ORPPT.

5 Conclusion

Within this paper, an end-to-end optimization is proposed to formulate fusion and detection in a harmonious one-stage training process. We introduce a object-region-pixel phylogenetic tree structure and coarse-to-fine diffusion process to simulate these two tasks in different visual perceptions needed for diverse task requirements. In addition, we align the orthogonal components of the fusion detection linear system of the gradients by gradient matrix task-alignment. By unrolling the model to a welldesigned fusion network and diffusion-based detection network, we can generate a visual-friendly result for fusion and object detection in an efficient way without tedious training steps and inherent optimization barriers.

[1] Kai Jiang, Jiaxing Huang, Weiying Xie, Jie Lei, Yunsong Li, Ling Shao, and Shijian Lu. DABEV: Unsupervised Domain Adaptation for Bird s Eye View Perception, page 322 341. Springer Nature Switzerland, October 2024. ISBN 9783031730078. doi: 10.1007/978-3-031-73007-8_ 19. URL http://dx.doi.org/10.1007/978-3-031-73007-8_19.

[2] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 958 979, 2024.

[3] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning. IEEE Transactions on Circuits and Systems for Video Technology, 32(10):6700 6713, 2022.

[4] Wujie Zhou, Xinyang Lin, Jingsheng Lei, Lu Yu, and Jenq-Neng Hwang. Mffenet: Multiscale feature fusion and enhancement network for rgb thermal urban road scene parsing. IEEE Transactions on Multimedia, 24:2526 2538, 2021.

[5] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: Denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8082 8093, October 2023.

[6] Wujie Zhou, Shaohua Dong, Caie Xu, and Yaguan Qian. Edge-aware guidance fusion network for rgb thermal scene parsing. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 36, pages 3571 3579, 2022.

[7] Hui Li, Xiao-Jun Wu, and Josef Kittler. Rfn-nest: An end-to-end residual fusion network for infrared and visible images. Information Fusion, 73:72 86, 2021.

[8] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023.

[9] Jinyuan Liu, Xin Fan, Ji Jiang, Risheng Liu, and Zhongxuan Luo. Learning a deep multi-scale feature ensemble and an edge-attention guidance for image fusion. IEEE Transactions on Circuits and Systems for Video Technology, 32(1):105 119, 2021.

[10] Zhanbo Huang, Jinyuan Liu, Xin Fan, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Reconet: Recurrent correction network for fast and efficient multi-modality image fusion. In European Conference on Computer Vision (ECCV), pages 539 555. Springer, 2022.

[11] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In International Joint Conference on Artificial Intelligence (IJCAI), 2022.

[12] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multimodality image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5906 5916, 2023.

[13] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Jiangshe Zhang, and Pengfei Li. Didfuse: Deep image decomposition for infrared and visible image fusion. In International Joint Conference on Artificial Intelligence (IJCAI), pages 970 976. ijcai.org, 2020.

[14] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Detfusion: A detection-driven infrared and visible image fusion network. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4003 4011, 2022.

[15] Jinyuan Liu, Zhu Liu, Guanyao Wu, Long Ma, Risheng Liu, Wei Zhong, Zhongxuan Luo, and Xin Fan. Multi-interactive feature learning and a full-time multi-modality benchmark for image fusion and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8115 8124, 2023.

[16] Linfeng Tang, Jiteng Yuan, and Jiayi Ma. Image fusion in the loop of high-level vision tasks: A semantic-aware real-time infrared and visible image fusion network. Information Fusion, 82: 28 42, 2022.

[17] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5802 5811, 2022.

[18] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via meta-feature embedding from object detection. In CVPR, pages 13955 13965, 2023.

[19] Hui Li, Xiao-Jun Wu, and Tariq Durrani. Nestfuse: An infrared and visible image fusion architecture based on nest connection and spatial/channel attention models. IEEE Transactions on Instrumentation and Measurement, 69(12):9645 9656, 2020.

[20] Risheng Liu, Jinyuan Liu, Zhiying Jiang, Xin Fan, and Zhongxuan Luo. A bilevel integrated model with data-driven layer ensemble for multi-modality image fusion. IEEE Transactions on Image Processing, 30:1261 1274, 2020.

[21] Jiayi Ma, Han Xu, Junjun Jiang, Xiaoguang Mei, and Xiao-Ping Zhang. Ddcgan: A dualdiscriminator conditional generative adversarial network for multi-resolution image fusion. IEEE Transactions on Image Processing, 29:4980 4995, 2020.

[22] Jiayi Ma, Wei Yu, Pengwei Liang, Chang Li, and Junjun Jiang. Fusiongan: A generative adversarial network for infrared and visible image fusion. Information Fusion, 48:11 26, 2019.

[23] Jiayi Ma, Hao Zhang, Zhenfeng Shao, Pengwei Liang, and Han Xu. Ganmcc: A generative adversarial network with multiclassification constraints for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 70:1 14, 2020.

[24] Nirmala Paramanandham and Kishore Rajendiran. Infrared and visible image fusion using discrete cosine transform and swarm intelligence for surveillance applications. Infrared Physics & Technology, 88:13 22, 2018.

[25] Risheng Liu, Zhu Liu, Jinyuan Liu, and Xin Fan. Searching a hierarchically aggregated fusion architecture for fast multi-modality image fusion. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1600 1608, 2021.

[26] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502 518, 2020.

[27] Zhu Liu, Jinyuan Liu, Guanyao Wu, Long Ma, Xin Fan, and Risheng Liu. Bi-level dynamic learning for jointly multi-modality image fusion and beyond. ar Xiv preprint ar Xiv:2305.06720, 2023.

[28] Xin Zhang, Weiying Xie, Yunsong Li, Jie Lei, Kai Jiang, Leyuan Fang, and Qian Du. Blockwise partner learning for model compression. IEEE Transactions on Neural Networks and Learning Systems, pages 1 14, 2023. doi: 10.1109/TNNLS.2023.3306512.

[29] Kai Jiang, Jiaxing Huang, Weiying Xie, Jie Lei, Yunsong Li, Ling Shao, and Shijian Lu. Domain adaptation for large-vocabulary object detectors, 2024. URL https://arxiv.org/abs/2401. 06969.

[30] Maoxun Yuan, Yinyan Wang, and Xingxing Wei. Translation, scale and rotation: Cross-modal alignment meets rgb-infrared vehicle detection. In European Conference on Computer Vision (ECCV), pages 509 525. Springer, 2022.

[31] Xue Yang, Jirui Yang, Junchi Yan, Yue Zhang, Tengfei Zhang, Zhi Guo, Xian Sun, and Kun Fu. Scrdet: Towards more robust detection for small, cluttered and rotated objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8232 8241, 2019.

[32] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 35, pages 3163 3171, 2021.

[33] Xue Yang, Junchi Yan, Wenlong Liao, Xiaokang Yang, Jin Tang, and Tao He. Scrdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rotation loss smoothing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2): 2384 2399, 2022.

[34] Xue Yang, Gefan Zhang, Xiaojiang Yang, Yue Zhou, Wentao Wang, Jin Tang, Tao He, and Junchi Yan. Detecting rotated objects as gaussian distributions and its 3-d generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4335 4354, 2022.

[35] R Caruana. Multitask learning: A knowledge-based source of inductive bias1. In International Conference on Machine Learning (ICML), pages 41 48. Citeseer, 1993.

[36] Sebastian Ruder. An overview of multi-task learning in deep neural networks. ar Xiv preprint ar Xiv:1706.05098, 2017.

[37] Felix JS Bragman, Ryutaro Tanno, Sebastien Ourselin, Daniel C Alexander, and Jorge Cardoso. Stochastic filter groups for multi-task cnns: Learning specialist and generalist convolution kernels. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1385 1394, 2019.

[38] Iasonas Kokkinos. Ubernet: Training a universal convolutional neural network for low-, mid- , and high-level vision using diverse datasets and limited memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6129 6138, 2017.

[39] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. ar Xiv preprint ar Xiv:1904.02920, 2019.

[40] Chanho Ahn, Eunwoo Kim, and Songhwai Oh. Deep elastic networks with model selection for multi-task learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6529 6538, 2019.

[41] Ximeng Sun, Rameswar Panda, Rogerio Feris, and Kate Saenko. Adashare: Learning what to share for efficient deep multi-task learning. Advances in Neural Information Processing Systems (Neur IPS), 33:8728 8740, 2020.

[42] Clemens Rosenbaum, Tim Klinger, and Matthew Riemer. Routing networks: Adaptive selection of non-linear functions for multi-task learning. International Conference on Learning Representations (ICLR), 2017.

[43] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7482 7491, 2018.

[44] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International Conference on Machine Learning (ICML), pages 794 803. PMLR, 2018.

[45] Minh-Thang Luong, Quoc V Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task sequence to sequence learning. International Conference on Learning Representations (ICLR), 2015.

[46] Jean-Antoine Désidéri. Mutiple-gradient descent algorithm for multiobjective optimization. In European Congress on Computational Methods in Applied Sciences and Engineering (ECCOMAS 2012), 2012.

[47] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. Advances in Neural Information Processing Systems (Neur IPS), 34: 18878 18890, 2021.

[48] Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Towards impartial multi-task learning. In International Conference on Learning Representations (ICLR), 2021.

[49] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multi-task learning as a bargaining game. ar Xiv preprint ar Xiv:2202.01017, 2022.

[50] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems (Neur IPS), 33:5824 5836, 2020.

[51] Dmitry Senushkin, Nikolay Patakin, Arseny Kuznetsov, and Anton Konushin. Independent component alignment for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20083 20093, 2023.

[52] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19830 19843, 2023.

[53] Alexander Toet. The tno multiband image data collection. Data in Brief, 15:249 251, 2017.

[54] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2.

https://github.com/facebookresearch/detectron2, 2019.

[55] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. ar Xiv preprint ar Xiv:1906.07155, 2019.

[56] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, Wenwei Zhang, and Kai Chen. Mmrotate: A rotated object detection benchmark using pytorch. In Proceedings of the 30th ACM International Conference on Multimedia, 2022.

[57] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16794 16805, 2023.

[58] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83: 79 92, 2022.

[59] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swin transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200 1217, 2022.

[60] ultralytics/yolov5: v3.1 - Bug Fixes and Performance Improvements. https://doi.org/10. 5281/zenodo.4154370, 2020.

[61] Fang Qingyun, Han Dapeng, and Wang Zhaokui. Cross-modality fusion transformer for multispectral object detection. ar Xiv preprint ar Xiv:2111.00273, 2021.

[62] Jifeng Shen, Yifei Chen, Yue Liu, Xin Zuo, Heng Fan, and Wankou Yang. Icafusion: Iterative cross-attention guided feature fusion for multispectral object detection. Pattern Recognition, 145:109913, 2024.

[63] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980 2988, 2017.

[64] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (Neur IPS), 28, 2015.

[65] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on the horizontal bounding box for multi-oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(4):1452 1459, 2020.

[66] Xiao He, Chang Tang, Xin Zou, and Wei Zhang. Multispectral object detection via crossmodal conflict-aware learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1465 1474, 2023.

[67] J Wesley Roberts, Jan A Van Aardt, and Fethi Babikker Ahmed. Assessment of image fusion procedures using entropy, image quality, and multispectral classification. Journal of Applied Remote Sensing, 2(1):023522, 2008.

[68] Guihong Qu, Dali Zhang, and Pingfan Yan. Information measure for performance of image fusion. Electronics Letters, 38(7):1, 2002.

[69] Yu Han, Yunze Cai, Yin Cao, and Xiaoming Xu. A new image fusion performance metric based on visual information fidelity. Information Fusion, 14(2):127 135, 2013.

[70] Liqiang He and Sinisa Todorovic. Destr: Object detection with split transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9377 9386, 2022.

Three metrics are used for MF evaluation: entropy (EN) [67], mutual information (MI) [68] and visual information fidelity (VIF) [69]. EN evaluates the information richness in an image, and the higher EN means more information. MI evaluates the information similarity between the input images and fused images. The higher MI illustrates more information of the input images is fused. VIF measures the ability to extract visible information from the input image, and a larger VIF represents less visible distortion in the fused result. Here, we use the V channel in the HSV space of the fusion results to calculate these metrics. Moreover, we use m AP50:95 [70] to comprehensively evaluate OD performance, where the average of m APs sampling every 5 from AP50 to AP95 is calculated. A higher m AP50:95 means a better OD effect.

A.2 Experiment setting on YOLOv5s

To more effectively evaluate the fusion images and observe their impact on downstream detection tasks, we conduct tests using the baseline detector YOLOv5s on all SOTA methods on the M3FD dataset. All images are resized to 1024 1024 and trained from scratch for 300 epochs with a batch size of 64. In addition, the experiments for Drone Vehicle dataset are conducted on YOLOv5s-OBB using the generated fusion images, all images are resized to 640 640 and trained from scratch for 50 epochs with a batch size of 64.

A.3 Experiments on Drone Vehicle

Drone Vehicle comprises aerial RGB-IR images captured by drones, encompassing various scenes from an aerial perspective. It contains five categories of target objects. The dataset consists of 28,439 pairs of images, divided into training, validation, and test sets. We train the MF network on the training set and perform inference directly on the test set. Finally, fusion images are generated for both the training and test sets. All detection accuracies are obtained on the test set. The visualization results of the fusion and detection of Drone Vehicle are shown in Figure 7 and 8. In the process of generating fusion images, we first obtain RGB images, then convert them to HSV format, and only take the V channel as the final fusion image. It preserves the brightness information of the objects while containing rich details from visible images. As shown in Figure 7, the first row and the second row are images extracted from the training set and the test set, respectively. Although the style of the fusion images resembles that of the infrared images, by combining visible image information, the fusion images can differentiate different parts of each object, reflecting varying degrees of brightness information. This is advantageous for fine-grained classification. Additionally, as shown in Figure 8, the visually friendly fusion images also assist in the object detection task, allowing for sufficient detection even in cases of small and dense objects.

Figure 7: Qualitative results of image fusion on Drone Vehicle.

A.4 Experiments on Coase-to-Fine Diffusion Process

We conducted ablation experiments on CFDP in Table 8, investigating its inclusion and the number of proposed boxes. In the setting without CFDP, we maintained the backbone network while substituting

Figure 8: Visual results of object detection on Drone Vehicle.

CFDP with RPN (Region Proposal Network), standard components in two-stage object detectors. Results indicate that CFDP enhances detailed information capture and precise box guidance, thereby enhancing fusion image quality and detection performance. For optimal balance between performance and efficiency, we selected 500 proposal boxes.

Table 8: Ablation studies of the CFDP on M3FD dataset. Settings Proposal boxes EN MI VIF m AP50 m AP50:95 Tr.Time w/o CFDP 500 5.71 14.39 1.45 90.13 61.98 2h52m11s

300 6.01 14.57 1.53 90.89 63.29 2h23m45s 500 6.36 15.47 1.65 91.80 63.83 2h50m32s 1000 6.37 15.34 1.63 92.05 63.75 3h32m30s

A.5 More Fusion Visualization Results

More comparisons of infrared-visible image fusion visualization results are depicted in Figure 9, 10 and 11. These fusion results demonstrate the advantages of synchronously optimizing fusion and detection tasks. With minimal training costs, we obtain fusion images that are visually and detection-friendly. Specifically, the fusion images retain significant object information extracted from infrared images, while also preserving detailed information such as texture, color, and background from visible images. Our method effectively combines the strengths of both modalities to enhance the overall performance of the detection.

Figure 9: Qualitative results of image fusion on M3FD.

A.6 More Detection Visualization Results

The visualization of infrared-visible object detection is demonstrated in Figure 12. Our fusion images are noticeably superior due to the effective combination of object information from infrared images and texture details from visible images. This integration enables the object detection network to distinguish between the background and the object clearly. Additionally, for small and occluded objects, the clear edge details assist the network in achieving improved detection results.

A.7 Limitations and Broader Impacts

Limitations. Our E2E-MFD approach is effective for the joint learning of the multimodal fusion and object detection task and has been validated on various datasets. However, the validations of the current model rely on the visible and infrared modalities. Constrained by the scarcity of relevant datasets within the community, the paper lacks validation with additional datasets containing new modalities. Future research will focus on addressing this gap by exploring, constructing, and incorporating more diverse multimodal datasets serving multimodal fusion detection.

Broader Impacts. Our paper aims to broaden the applicability of joint learning of multimodal data and object detection to various research domains. However, this broader scope may present challenges when using the model in domains that include harmful content. These challenges arise from the data itself, rather than from the model. Therefore, it is crucial to have adequate data regularization to effectively address these concerns.

Figure 10: Qualitative results of image fusion on Road Scene.

Figure 11: Qualitative results of image fusion on TNO.

Figure 12: Visual results of object detection on M3FD.

Neur IPS Paper Checklist

Question: Do the main claims made in the abstract and introduction accurately reflect the paper s contributions and scope?

Answer: [Yes] Justification: The abstract and introduction accurately state the contributions of the E2EMFD model, including its novel design of synchronous joint learning of the multimodal fusion and object detection.

Guidelines:

The answer NA means that the abstract and introduction do not include the claims made in the paper. The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.

2. Limitations

Question: Does the paper discuss the limitations of the work performed by the authors?

Answer: [Yes]

Justification: The limitation of the the work is detailed discussed in the Appendix A.7.

Guidelines:

The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. The authors are encouraged to create a separate "Limitations" section in their paper. The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.

3. Theory Assumptions and Proofs

Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?

Answer: [NA] Justification: NA. Guidelines:

The answer NA means that the paper does not include theoretical results. All the theorems, formulas, and proofs in the paper should be numbered and crossreferenced. All assumptions should be clearly stated or referenced in the statement of any theorems. The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental Result Reproducibility

Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: All the necessary information needed to reproduce the main experimental results are fully stated in the Section 4 and Appendix A. This information ensures the understanding the results which support the main claims and conclusions of the paper. Guidelines:

The answer NA means that the paper does not include experiments. If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. While Neur IPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code

Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The paper provide the open access to the data and code with sufficient instructions for reproducing the main experimental results in the supplementary material. Guidelines:

The answer NA means that paper does not include experiments requiring code. Please see the Neur IPS code and data submission guidelines (https://nips.cc/ public/guides/Code Submission Policy) for more details. While we encourage the release of code and data, we understand that this might not be possible, so No is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). The instructions should contain the exact command and environment needed to run to reproduce the results. See the Neur IPS code and data submission guidelines (https: //nips.cc/public/guides/Code Submission Policy) for more details. The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental Setting/Details

Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes]

Justification: The training and test details including data splits, hyperparameters and type of optimizer are illustrated in Section 4.1 and Appendix A.3 to ensure the results of the paper can be fully understand and reproduced. Guidelines:

The answer NA means that the paper does not include experiments. The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment Statistical Significance

Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes]

Justification: The appropriate information about the statistical significance of the experiments are carefully reported in Section 4 ensuring that the findings are presented with an appropriate level of precision and confidence. Guidelines:

The answer NA means that the paper does not include experiments. The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.

The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) The assumptions made should be given (e.g., Normally distributed errors). It should be clear whether the error bar is the standard deviation or the standard error of the mean. It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments Compute Resources

Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The paper provide the sufficient information on the computation computer resources including the type of compute works (Ge Force RTX 3090 GPU), memory, and time of execution in the Section 4. Guidelines:

The answer NA means that the paper does not include experiments. The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn t make it into the paper). 9. Code Of Ethics

Question: Does the research conducted in the paper conform, in every respect, with the Neur IPS Code of Ethics https://neurips.cc/public/Ethics Guidelines? Answer: [Yes] Justification: The research conducted in the paper conforms with the Neur IPS Code of Ethics which ensuring all aspects of our study, including data collection, analysis, and reporting. Guidelines:

The answer NA means that the authors have not reviewed the Neur IPS Code of Ethics. If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader Impacts

Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: Broader Impacts are discussed in paper. We highlight the benefits of our research, such as advancing the field of multimodal fusion detection learning, which can lead to improvements in various applications, including autonomous vehicles and remote sensing. At the same time, the limitations of the method are analysed in Appendix A.7.

Guidelines:

The answer NA means that there is no societal impact of the work performed. If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards

Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: NA. Guidelines:

The answer NA means that the paper poses no such risks. Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 12. Licenses for existing assets

Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: The creators or original owners of assets are properly credited and the license and terms of use are explicitly mentioned and properly respected. Guidelines:

The answer NA means that the paper does not use existing assets. The authors should cite the original paper that produced the code package or dataset. The authors should state which version of the asset is used and, if possible, include a URL. The name of the license (e.g., CC-BY 4.0) should be included for each asset.

For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. If this information is not available online, the authors are encouraged to reach out to the asset s creators.

13. New Assets

Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?

Answer: [Yes]

Justification: We will open the code totally if the paper is accepted with the new assets documented and documentation is provided alongside the assets.

Guidelines:

The answer NA means that the paper does not release new assets. Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. The paper should discuss whether and how consent was obtained from people whose asset is used. At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.

14. Crowdsourcing and Research with Human Subjects

Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?

Answer: [NA]

Justification: The paper does not involve crowdsourcing nor research with human subjects.

Guidelines:

The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. According to the Neur IPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.

15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects

Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?

Answer: [NA]

Justification: The paper does not involve crowdsourcing nor research with human subjects..

Guidelines:

The answer NA means that the paper does not involve crowdsourcing nor research with human subjects.

Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the Neur IPS Code of Ethics and the guidelines for their institution. For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.