# learning_to_discover_and_detect_objects__5de61f09.pdf

Learning to Discover and Detect Objects

Vladimir Fomenko1 Ismail Elezi2 Deva Ramanan3

Laura Leal-Taixé2 Aljoša Ošep2,3

1Microsoft Azure AI 2Technical University of Munich 3Carnegie Mellon University vladfomenko@microsoft.com {ismail.elezi, leal.taixe, aljosa.osep}@tum.de deva@cs.cmu.edu

We tackle the problem of novel class discovery and localization (NCDL). In this setting, we assume a source dataset with supervision for only some object classes. Instances of other classes need to be discovered, classified, and localized automatically based on visual similarity without any human supervision. To tackle NCDL, we propose a two-stage object detection network Region-based NCDL (RNCDL) that uses a region proposal network to localize regions of interest (Ro Is). We then train our network to learn to classify each Ro I, either as one of the known classes, seen in the source dataset, or one of the novel classes, with a long-tail distribution constraint on the class assignments, reflecting the natural frequency of classes in the real world. By training our detection network with this objective in an end-to-end manner, it learns to classify all region proposals for a large variety of classes, including those not part of the labeled object class vocabulary. Our experiments conducted using COCO and LVIS datasets reveal that our method is significantly more effective than multi-stage pipelines that rely on traditional clustering algorithms. Furthermore, we demonstrate the generality of our approach by applying our method to a large-scale Visual Genome dataset, where our network successfully learns to detect various semantic classes without direct supervision. Our code is available at https://github.com/vlfom/RNCDL.

1 Introduction

We tackle novel class discovery and localization in unlabeled datasets, a long-standing problem in computer vision [55, 59, 41, 42, 54]. As we are approaching the limits of well-understood and successful supervised training of object detectors [22, 21, 53, 27] and labeling novel and rare classes that appear in the long tail of object class distribution is becoming prohibitively expensive [24], we expect joint novel class discovery and localization to become increasingly more important.

Object class discovery in image collections is a difficult problem, as there are, in general, many possible equally valid attributes (e.g., object color or orientation) for grouping of visual patterns. As shown in [29, 30], novel class discovery (NCD) can be well-posed and tackled in a data-driven manner by injecting prior knowledge on how we wish to group semantic classes using some degree of supervision. However, existing NCD methods [29, 30, 25, 26, 72, 7, 73, 74, 19] all assume curated datasets, where objects of interest are pre-cropped and semantic classes are fairly balanced. By contrast, in this paper, we tackle joint novel class discovery and localization (NCDL), a task of learning to discover and detect objects from raw, unlabeled, and uncurated data. This is a significantly more challenging and realistic problem setting, as images always contain a mixture of labeled and unlabeled object classes, and we need to localize and categorize them.

Work done while at Technical University of Munich.

36th Conference on Neural Information Processing Systems (Neur IPS 2022).

Figure 1: Novel class discovery and localization. Given labeled images along with annotations for known, frequently observed semantic classes and unlabeled images that may contain instances of novel classes, our network learns to localize and recognize common semantic classes and categorize novel classes, for which supervision in the form of labeled instances is not available.

We re-purpose COCO [43] and LVIS [24] datasets to study this problem. We use 50% of COCO images together with COCO class vocabulary and labels for supervised model training. We treat the remaining 50% of images as unlabelled image collection. We use these images to learn to detect remaining object classes for which supervision is not given. With this setting, we mimic real-world scenarios where no supervision is given for the target domain.

We train our two-stage [53, 27] object detector Region-based NCDL (RNCDL) in an end-to-end manner such that we can (i) correctly detect (i.e., localize and categorize) labeled objects, and, in parallel, (ii) detect unlabeled objects by learning feature representations suitable for categorization of objects that would otherwise be classified as the background class. We start with supervised training on the labeled source domain, followed by self-supervised training on the target (unlabeled) domain. Finally, we transfer knowledge from the source to the target domain by retaining weights for class-agnostic modules, such as RPN, bounding box regression, and instance segmentation heads. For classification, we add to the primary known classification head a new, secondary head for categorization of regions that do not correspond to these a-priori known semantic classes. We train both heads jointly with an objective to categorize every region proposal consistently under multiple transformations and a constraint that the marginal distribution of class assignments should follow a long-tailed distribution. Such a non-uniform classification prior incentivizes the network to learn features that capture the diversity of objects and prevents the network from being biased towards the labeled, known semantic classes.

Our end-to-end trainable RNCDL network categorizes all regions of interest in a single network forward pass and does not require running clustering algorithms during the inference. Moreover, it significantly outperforms prior art [68] that learns object representations using contrastive learning followed by k-means clustering. More importantly, with this work, we lay the foundations for an exciting new line of research beyond well-established supervised learning for object detection.

In summary, this paper makes the following contributions: (i) we study the problem of novel class discovery and localization in-the-wild using standard object detection datasets that do not pre-localize regions of interest manually or assume objects are roughly centered, i.e., exhibit a strong photographer bias. (ii) we propose a simple yet effective and scalable method that can be trained end-to-end on the unlabeled domain. First, our network is bootstrapped using a labeled dataset to learn features suitable for categorizing future instances. Then, we learn feature representations suitable for categorization for arbitrary objects using the self-supervised objective. Finally, (iii), we show that our method can discover and detect novel class instances and outperform prior work, demonstrating the generalization to datasets beyond COCO.

2 Related Work

Object discovery in image and video collections. The discovery of semantic classes that appear in the unlabeled image or video collections is a long-standing research problem [60]. Early methods

employ multiple bottom-up segmentation of images [55, 59] to discover commonly occurring patterns using topic modeling, e.g., Dirichlet allocation [6].

Method by [41] relies on saliency, while [42] incorporates semantic knowledge about certain semantic classes for which supervision is available. These methods are evaluated on smaller datasets [23, 58, 17] that contain one or a handful of well-delineated objects. Rubinstein et al. [54] tackles object discovery methods in nosy internet photo collections. Beyond images, Kwak et al. [40] tackle the discovery of objects in You Tube videos that usually contain a single object exhibiting dominant motion. They build on a two-stage approach: (i) identify the most salient region proposals in terms of appearance and motion, and (ii) co-segment consistently appearing objects across video clips. Object discovery in videos, recorded from a moving platform, was tackled in [49, 50] by first mining video-object tracks by associating region proposal network (RPN) [53] based proposals across time, followed by clustering of video proposals based on features, extracted from a pre-trained network. Recent work on open-set panoptic segmentation [32] similarly groups object proposals based on features extracted from a pre-trained network to pseudo-label commonly-occurring objects. The method by [68] follows the general approach of clustering RPN proposals; however, instead of relying on pre-trained CNN features, this approach learns representations suitable for clustering using self-supervised objective, trained via contrastive learning.

While most of the foregoing works focus on saliency criteria to significantly narrow down the number of region proposals, [62, 63] resort to combinatorial optimization that can cope with a large set of object proposals to discover objects that frequently appear in image datasets. Furthermore, it has been shown [64] that this approach can be scaled to large image collections, e.g., Open Images [39].

Novel class discovery. Novel class discovery (NCD) methods assume labeled data for some semantic classes is given. This data can be used as a guide to learn a notion of similarity and invariances to certain features that should be ignored for the grouping (such as object color and orientation). Unlabeled images are assumed to contain novel class instances only. Hsu et al. [29, 30] propose a framework that injects prior knowledge on how we wish to group semantic classes from supervised learning on labeled images to the unlabeled domain via learnable similarity function. The work of Han et al. [25] builds on deep clustering [71] and suggests an approach for automatically discovering the number of novel classes. Han et al. [26] build on the pairwise similarity of image features [12] and propose leveraging self-supervised pretraining along with freezing the backbone to reduce bias towards the known classes. Follow-up works [72, 34, 74, 73] further boost the performance by leveraging self-supervised learning and regularization techniques. UNO [19] proposes a simple method based on Sw AV [11] that bootstraps signal for novel class training using pseudo-labels generated under equipartitioning constraints. ORCA [7] extends the NCD problem setting and assumes that we are given images containing instances of novel and known objects at inference time. Motivated by [61, 12, 65], they leverage self-supervised initialization of backbone, supervised loss, and unsupervised loss based on pairwise pseudo-labels, and novel dynamic softmax margin [65] to alleviate bias of the known classes. By contrast to our work, the aforementioned NCD methods investigate the problem in the image classification setting, assuming objects of interest are localized (pre-cropped) and operate in the absence of outliers (e.g., images containing no objects).

Related and complementary research directions. Methods for generic [45] or large vocabulary object detection [24] focus on detecting/segmenting a large set of a-priori known semantic classes, often by utilizing multiple datasets and (possibly weak) supervision [75, 52, 31]. A special case is zero-shot learning [70] and detection [3, 48, 51], where (mostly weak) supervision for unseen classes comes in the form of attributes, class names, latent features, or other forms of auxiliary data commonly used to learn a joint multi-modal embedding space. Few-shot learning methods [36, 67, 18, 8] learn from one or a few data samples per class. Methods by [18, 8] employ a secondary classification head that is trained using few-shot supervision and distillation [18] or margin loss [8]. In contrast to zero-shot, few-shot, and large-vocabulary detection, we assume no supervision and prior knowledge for object classes that appear in the tail.

Methods for open-set recognition [56, 57, 33, 5] and detection [48, 16] focus on the calibration of per-class uncertainties to minimize the confusion between known and possibly unknown classes. As defined by [4, 46], open-world recognition methods must explicitly recognize unknown object instances that were not observed as labeled samples during the model training and must continually update object detectors to recognize these unknown instances. Recent work on open-world detection [35] focuses on the continual learning aspect of the task. For the evaluation, labels for novel

Figure 2: A high-level overview of our network. (top) During the supervised training phase, we train our backbone and RPN networks using labeled data, together with classification head and a class agnostic localization head. During the discovery phase, we freeze all the layers of the network apart from classification head and attach and train a novel classification head using unlabeled data. During the inference (bottom), we perform a standard R-CNN pass, using classification heads of both known and novel categories to predict a class assignment for each proposal. This can be either one of K classes, that were presented a labeled samples during the model training, or any novel object class that appears in the training data.

classes are fully provided (those are the held-out classes). Our work is orthogonal and tackles the missing bit of the open-world detection pipeline: semantic grouping of novel objects that can appear in the training data. Clusters of discovered novel classes could be verified by human annotators and used to update detection models continually, as proposed by [4].

3 Novel Class Discovery and Localization

We first detail novel class discovery and localization (NCDL), the task of learning to discover and detect objects from unlabeled and uncurated data. We assume given a source (labeled) image dataset Dk for known object classes Ck (|Ck| = K), that provides both class and box level objects supervision: Dk = {(I, {(b, c)})}, where I Rh w 3, b R4, and c Ck. In addition, we use one or more target (unlabeled) datasets Dn = {I} that may contain instances of both previously seen and novel object classes Ck Cn (|Cn| = N, N > K) that we need to learn to detect in the absence of any labels. For an arbitrary image I from the source or target domain, NCDL methods should detect all object instances present in the image: {(b , c )}, i.e., localize them and categorize them. For example, an instance of a known class, e.g., car, needs to be detected as the car class, while a novel class instance with c Cn needs to be assigned to one of the unknown, novel classes (e.g., unknown_1, unknown_2, etc.).

As this paper s main contribution, we propose Region-based NCDL (RNCDL), an end-to-end trainable detection and discovery network that learns to localize and categorize objects, including those for which explicit supervision is not given. We base our method on a two-stage Faster or Mask RCNN [53, 27], trained in a supervised manner on a dataset that contains labels for some semantic classes. Such detectors can already reliably localize and classify known objects and are naturally suitable for the NCDL task, as their region proposal network (RPN) is trained in a class-agnostic manner to propose regions of interest (Ro Is) that likely contain objects. Given object-centric feature representation for each localized Ro I, one natural approach for discovering (categorizing) novel classes would be to cluster such regions in feature space using, e.g., k-means, c.f., [49, 32, 68]. Feature representations learned via supervised R-CNN training are suitable for localizing regions that likely contain objects and for classifying them as one of the known classes. However, features

representing Ro Is that do not contain labeled class objects are effectively trained to be collapsed into a single background class and, therefore, not discriminative enough for robust categorization of novel instances. Furthermore, such an approach would have a strong bias toward classifying novel instances as one of the labeled classes.

3.1 Method Overview

In the following, we detail our RNCDL, which learns to categorize all region proposals jointly on both (labeled) source and (unlabeled) target datasets in an end-to-end manner. In a nutshell, we first train class-agnostic modules of our detector on the source domain (Fig. 2A, top), i.e., RPN, bounding box regression head, (optionally) instance segmentation head, and a classifier for known classes. We proceed with self-supervised training (Fig. 2A, bottom) on the target domain to learn to additionally categorize instances of unknown classes. To this end, we add a secondary classification head (to the primary classification head, trained to categorize known classes) and train both jointly, with an objective to categorize every region proposal consistently under multiple transformations and a constraint that the marginal distribution of class assignments follows a long-tailed distribution.

Supervised training on the source domain. We begin with a self-supervised backbone pretraining [28] on Image Net [15] to initialize the R-CNN backbone network. We then train the object detector on the source (labeled) domain using supervision for objects of known Ck classes, such as 80 COCO classes (Fig. 2A, top). We use the labeled data to train the shared feature extractor (backbone) together with a class-agnostic region proposal network (RPN), class-agnostic box regression head, and a classification head for the categorization of Ck classes. Specifically, we train RPN to predict a set of class-agnostic proposals {(b, o)}, for each image, where b R4 are predicted coordinates and o R is the objectness score. We then use Ro IAlign [27] to crop Ro I features from the backbone and feed them to an MLP to obtain box-head features ˆf R ˆ F that we pass to the additional network heads to obtain the final predictions. We train a class-agnostic bounding box regression head to refine box location for each Ro I ˆb = hbb(b,ˆf), and a classification box head to categorize each Ro I as one of the K classes or the background class: ˆc = hk(ˆf) = Wkˆf, where ˆc RK+1 and Wk R(K+1) ˆ F . We train our network using the standard R-CNN loss [53] Lsup = LRP N +Lbox +Lcls, that consists of RPN classification LRP N loss, box regression loss Lbox, and second-stage classification loss Lcls. We discuss all changes to Detectron2 Faster R-CNN implementation, as needed for the NCDL task, in Sec. ?? in the appendix.

Discovery on the target domain. For the self-supervised (discovery) training phase on the target domain, we first modify the primary (known) classification head by removing the background class from its weight matrix such that ˆ Wk RK ˆ F . Then, we attach a secondary (novel) classification head hn on top of Ro I box features ˆf (Fig. 2A). This head consists of an MLP and a linear classification layer that we detail in Sec. 3.2.

In this phase we freeze all layers except the classification heads. We train both classification heads jointly using a self-supervised classification loss Lss (detailed in Sec. 3.3) and supervised classification loss Lcls: Ldisc = Lss + α Lcls where α is the supervised loss scale coefficient. We show in our ablations (Sec. 4.2) that keeping the downscaled (α < 1) supervised classification loss in the loop effectively alleviates forgetting the weights for the known classes, while large α may induce an unwanted bias towards classifying Ro Is as known classes.

Both classification losses are effectively cross-entropy losses. Lcls is trained using provided labels, while Lss is trained using pseudo-labels generated online for the current batch of proposals. We detail online pseudo-label generation in Sec. 3.3. In a nutshell, this loss encourages the network to categorize every region proposal consistently under multiple transformations and, importantly, under a constraint that the marginal distribution of predicted class assignments for the last batches follows a prior long-tailed distribution. We compute the supervised loss only for annotation-matched region proposals. With such a joint training objective, our network is encouraged to learn to distinguish a large variety of classes present in the regions sampled by RPN while retaining the capability to classify known objects.

3.2 Secondary Classification Head

As the box features ˆf are kept frozen during the discovery phase, we use several non-linear projection layers in the secondary classification head: ˆf n = gn(ˆf), where gn is a multi-layer perceptron with

Re LU activations. With this transformation, we learn to disentangle features for distinct novel classes without affecting the Ro I box features ˆf. On top of the projected features ˆf n, we employ a cosine classification layer, shown to be effective in the context of few-shot learning [20, 13, 66]: ln = Wn ||Wn|| ˆf n

||ˆf n||, where ln RN are the predicted novel-class logits. We determine the number of novel classes (the parameter N for the secondary head) empirically (see Sec. 4.2).

3.3 Self-Supervision via Online Constrained Clustering

For self-supervised learning, we perform pseudo-labeling via constrained clustering during online model training to assign a soft pseudo-label to each Ro I (RPN proposal). We generate pseudo-labels q R(N+K) B for Ro Is in all input images (labeled and unlabeled, DN DK) that can span all classes, both known and novel. The parameter B denotes the batch size. We use these pseudo-labels q as a supervisory signal to compute cross-entropy loss Lss to train both classification heads.

For pseudo-label generation one possibility would be to use standard clustering methods (e.g., k-means [47]), however, such offline clustering would make training slow, and such methods were shown to produce unstable results in domain of image classification [9, 10]. Instead, we follow recent developments in the field of self-supervised representation learning [11, 2, 1] and minimize a clustering energy function online during the mode training (i.e., cross-entropy loss): E(ˆp, q) = 1 B PB i=1 PN+K y q(y|Ii) log ˆp(y|Ii), where q are class labels (i.e., pseudo-labels), and ˆp = softmax([lk, ln]) R(N+K) B are predicted class-probabilities. Note that both labels q and probabilities ˆp are subject to optimization. Similar to [2], we employ an alternating algorithm. First, we use a constrained clustering method to generate pseudo-labels q based on current network weights. Then, we update weights by optimizing ˆp using fixed q, and iterate.

To generate class pseudo-labels, we first choose a target clusters marginal probability distribution (i.e., a prior distribution). Empirically, we obtained best results using a non-symmetrical log-normal distribution. This distribution is right-skewed, suitable for modeling the long-tail. We then generate such soft pseudo-labels q that (i) closely match the chosen marginals and (ii) minimize E(ˆp, q) for the given ˆp. In [2, 1], it is shown that this can be posed as a constrained optimization problem and solved as an optimal transport problem using the fast online Sinkhorn-Knopp algorithm [14]. Having obtained the pseudo-labels q, we fix them, compute the loss Lss = E(ˆp, q), and update weights for the classification heads using the back-propagation algorithm.

Memory module. To generate good-quality pseudo-labels for each mini-batch, we need a diverse set of samples that capture a wide variety of classes. A small set of features per mini-batch could lead to noisy cluster assignments. As we operate in scenarios with more than a thousand classes to discover, we introduce a memory module (c.f., [28, 11]) to store Ro I box features from the last batches and ensure a diverse set of samples when computing clustering assignments. As in [28, 11], we design the memory module as a queue, where at the end of each iteration, we replace the oldest set of stored features with the features from the current batch. During pseudo-labeling, we pass the extended set of features to the Sinkhorn algorithm. Specifically, after extracting current-batch Ro I features, we concatenate them with the features stored in the memory module M, compute their logits, and as an input to the Sinkhorn algorithm use: ˆl = [ˆlk,ˆln] = [hk([ˆf, M]), hn([ˆf, M])], where M RMsz ˆ F and Msz is the number of features stored in the memory. Such a modified procedure generates B + Msz pseudo-labels for both current-batch proposals and memory features. We discard pseudo-labels generated for memory features and use only those B generated for the current-batch proposals, i.e., we use the memory module only to improve the quality of current-batch pseudo-labels.

Swapping multi-view cluster assignments. To ensure that the network outputs consistent classification predictions ˆp, we employ an additional supervisory signal from multiple augmentations [11]. Specifically, for each image from DN DK, we first generate Ro Is (proposals) and refine their coordinates via a class-agnostic classification head. We then generate two augmented views for each image (we discuss augmentations in Sec. ?? in the appendix), and extract features twice for each Ro I. Finally, we generate pseudo-labels for each view (set of features) individually, using the Sinkhorn algorithm as described in Sec. 3.3, obtaining two sets of pseudo-labels qi, i {1, 2} that correspond to the same proposals. To enforce feature invariance to the augmentations and capture the semantic similarity of the proposals, we swap the resulting pseudo-labels qi between views and calculate losses as E(ˆpi, qj), where i = j, i.e., we use pseudo-labels obtained for one view, as labels to the other. We

calculate the total loss as the average of the losses for both views: ˆLss = (E(ˆp1, q2)+E(ˆp2, q1))/2. For each view, we maintain a separate feature memory module.

3.4 Inference and Evaluation

Finally, to categorize Ro Is (RPN proposals) during the inference (Fig. 2B), we extract a set of Ro Is {(b, o)} for each image and pass them to localization refinement and class prediction heads. We pass Ro I features through both hk and hn classification heads and concatenate their logits: ˆp = softmax([lk, ln]) = softmax([hk(ˆf), hn(ˆf)]) RK+N to compute output class probabilities over K known and N novel semantic classes.

Semantic class assignment. Our method provides only categorization for novel classes. Class IDs (i.e., cluster IDs) are not directly mapped to classes, defined in a certain semantic class vocabulary, as needed for the evaluation. We therefore need to obtain a mapping of predicted class IDs to ground-truth semantic clsses. To do so, we follow similar strategy as prior art [25, 26, 19, 68]. Once we train our network, we generate classification predictions for each GT annotation in the validation dataset. Then, we use the Hungarian algorithm [38] to obtain an one-to-one mapping between GT labels and validation class predictions. We ignore instances of predicted classes that were not mapped to GT classes. Then, we proceed to the standard R-CNN inference pass, where we apply the generated mapping and follow the standard R-CNN post-processing and evaluation [43].

4 Experimental Evaluation

In this section, we discuss our evaluation test-bed, including datasets, evaluation settings, and metrics (Sec. 4.1). In Sec. 4.2 we justify our main design decisions by studying NCDL performance in a well-controlled setting that closely follows a real-world scenario (training on a source labeled and target unlabeled dataset). We then compare our method s performance to several baselines (Sec. 4.3) and, finally, demonstrate that our method is applicable beyond our evaluation test-bed (Sec. 4.4).

4.1 Evaluation Setting

COCOhalf + LVIS. We re-purpose COCO 2017 [43] and LVIS v1 [24] datasets for running ablations and comparisons with baselines and prior art. We use annotations for 80 COCO classes during the supervised training phase, aiming at further classifying additional 1000+ LVIS classes. COCO dataset contains 123K images with modal bounding box and segmentation mask annotations for 80 classes. The LVIS dataset contains a subset of 120K COCO images with annotations for 1203 classes that include all classes from COCO. We follow LVIS training and validation splits for our experiments, resulting in 100K training images and 20K validation images. In our NCDL setup, we further split 100K training images in half. We use only 50K images with 80-class annotations from COCO during the supervised training phase. We treat the rest of 50K images as additional unlabeled data used during self-supervised training (the discovery phase).

LVIS + Visual Genome. We perform a large-scale generalization experiment by using annotations for 1203 LVIS classes during the supervised learning (i.e., treating LVIS classes as labeled), aiming to learn to discover extra 2700+ classes from the Visual Genome (VG) v1.4 dataset [37] (i.e., treating VG classes as unlabeled). We provide dataset details and splits (that ensure LVIS and VG class vocabularies do not overlap) in the supplementary. We note that annotations provided in VG are not exhaustive per class, and many of its classes are abstract or semantically overlapping both within VG and when compared to LVIS. We thus mainly focus on qualitative results.

Implementation details. We base our model on Mask R-CNN [27] and FPN [44] implementations from Detectron2 [69]. We discuss implementation details in Sec. ?? of the appendix.

Evaluation metrics. For quantitative evaluation, we follow [25, 26, 19, 68] and match predicted cluster assignments with annotated semantic classes (Sec. 3.4). We follow common practice and report mean average precision (m AP@[.5:.95]) [43]. For qualitative results, we include instances of classes that were matched to annotated classes only. We detail our evaluation procedure in the supplementary.

Table 1: Impact of supervised loss strength. We check the results of our method as a function of the strength of the supervised loss.

Sup. loss coeff. m APall m APknown m APnovel 0 6.59 16.58 5.77 0.1 6.90 22.26 5.64 0.5 6.92 25.00 5.42 1.0 6.35 25.81 4.75

Table 2: Sensitivity to the number of classes. We vary the number of novel classes set during the discovery phase.

# novel classes m APall m APknown m APnovel 1000 5.42 23.24 3.95 3000 6.92 25.00 5.42 5000 6.24 24.91 4.70

Baselines. We compare our end-to-end trainable detector with prior work by Weng et al. [68], NCD methods ORCA [7], UNO[19], and k-means [47] baseline that all operate on cropped object proposals, provided by RPN. For NCD methods, we use labeled instance annotations from COCOhalf dataset as labeled images pool and cropped RPN proposals as unlabeled images pool. We extract proposals using the same R-CNN base network trained on COCOhalf to ensure comparable evaluation conditions. After training NCD classifiers, we use the same R-CNN to generate object proposals for the validation images, apply the classifiers on top of the proposals, and continue with the standard R-CNN post-processing steps. We provide more details in Sec. ?? in the appendix.

4.2 Model Ablations

We ablate single and multi-phase training, the impact of keeping supervised loss in the loop, the number of novel categories, pseudo-labeling prior, model architecture decisions, and backbone pretraining. In Sec. ?? in the appendix, we provide further results on experiments with hyperparameters for proposals extractor and how class-agnostic heads affect the performance of the supervised baseline.

Single-phase training and unfreezing R-CNN components during discovery. As shown in Table 3, we observe that training the network with both supervised and discovery losses from the beginning, summed up without any scaling, leads to divergence. We hypothesize that this is due to noisy gradients from our self-supervised signal, with R-CNN being very sensitive to the choice of losses and hyperparameters [53]. To further support this claim, we experiment with a standard supervised phase followed by a discovery phase without freezing any components. We observe that such a setup also leads to divergence. Unfreezing only the Ro I heads during the discovery phase does not lead to divergence, however, it reduces the score by 1.32 m AP. We thus keep all the components frozen during the discovery phase, apart from known and novel classification heads.

The strength of the supervised loss. We ablate the strength of the supervised classification loss that is used along with the discovery loss during the discovery phase. We vary the strength from 0 (no supervised loss) to 1 (no downscaling) and provide the results in Table 1. We observe that excluding supervised loss completely degrades the network s performance for the known classes and benefits the performance of the novel classes, while keeping the full supervised loss introduces a bias towards the known classes. This indicates a tradeoff between performance on known and novel classes. We set the strength of the supervised loss to 0.5 in all experiments.

Sensitivity to the number of novel classes. In Table 2 we experiment with varying the number of novel classes we use for the novel classification head. We observe that the best number of classes for the COCO + LVIS setup is 3000, while other choices result in lower scores. We conjecture that such scores are sensitive to the semantic classes defined by the target dataset with 3000 classes, the model learns the granularity and grouping that matches closest to the LVIS classes, but that does not necessarily mean that a finer or coarser granularity categorization is incorrect. In Sec. ?? in the appendix, we also experiment with attaching and training multiple heads jointly during the discovery phase and observe that scores slightly degrade in such setup but do not observe a significant drop in results.

Clustering prior. In Table 3, we experiment with using uniform prior distribution for class-marginals during pseudo-labels generation, as proposed in [2, 10]. We observe degradation of the score by 1.17 m AP, which confirms our hypothesis that a non-uniform log-normal prior distribution better models a real-world long-tailed class distribution. We use k-means for the online pseudo-labels generation as an additional ablation instead of constrained clustering. As a result, we observe a significant score degradation by 4.86 m AP, especially notable for the novel classes.

Table 3: Additional ablations. We experiment with removing individual components from the network and their impact on the overall performance.

Method m APall m APknown m APnovel RNCDL 6.92 25.00 5.42

Single-phase framework n/a n/a n/a Unfreeze all layers n/a n/a n/a Unfreeze Ro I heads 5.60 20.55 4.37

Uniform class prior 5.75 24.11 4.23 k-means [47] pseudo-lab. 2.06 21.13 0.48 W/o projection MLP 5.30 26.13 3.58 W/o swapped assignments 6.74 26.02 5.14 W/o memory 2.83 24.19 1.06 W/o Mo Co pretraining 4.77 16.92 3.77

Table 4: Comparison with state-of-the-art models. * as per open source code. adapted to support known classes in the target dataset. randomly initialized novel head.

Method m APall m APknown m APnovel Methods that operate on cropped proposals k-means [47] 1.33 15.61 0.14 Weng et al. [68] 1.62 17.85 0.27 ORCA [7] 2.03 20.57 0.49 UNO [19] 2.18 21.09 0.61

Methods that operate on FPN-based features k-means [47] 1.55 17.77 0.20 RNCDL w/ random init. 1.95 23.51 0.17 RNCDL 6.92 25.00 5.42

Fully-supervised 18.47 39.38 16.74

Embedding projector and swapped assignments. In Table 3, we demonstrate the effect of adding an MLP feature projector on top of Ro I box features for novel head classification. We observe that the performance of the network improves by 1.62 m AP. Furthermore, the same table shows that swapping pseudo-labels between proposals under multiple augmentations helps boost the score by 0.18 m AP.

Feature memory module. We demonstrate the importance of the memory module in Table 3. In all our experiments, we used the memory size of 100 batches unless stated otherwise. Removing the memory module significantly drops the performance by 4.09 m AP, especially for the novel classes.

Self-supervised pretraining matters. We compare the results of our framework under different backbone initialization employed for the supervised pretraining stage. For the randomly initialized backbone, we set all layers as trainable, while during the Mo Co initialization, the first two convolutional blocks are kept frozen (c.f., [28]). In Table 3 we observe that the Mo Co-initialized backbone improves the score by 2.15 m AP over randomly initialized backbone mode. We only pre-train our network using Image Net dataset [15], and not on the source or target object detection dataset.

Proposals extractor hyperparameters. We observe that when extracting the RPN proposals for self-supervised learning during the discovery phase, the absence of NMS yields the best results, and the optimal number of proposals per image is 50. We thus use such parameters for all the experiments.

4.3 Comparison to Prior Art

In Table 4, we compare our method with k-means baselines and three recent state-of-the-art methods that we adapted to our scenario. We present the results for both known (COCO) and unknown (the rest of LVIS) classes. Our method significantly outperforms the previous best NCD method UNO [19] by reaching 4.74 higher overall m AP. Furthermore, we outperform UNO in both the known and especially novel classes, where we reach 4.81 higher m AP. We also significantly outperform the approach by Weng et al. [68]. Our method is the only method to reach over 1 m AP in novel classes, where we reach 5.42 m AP. We believe the main advantages of our method to be the memory module, critical for effective self-supervision, and a backbone trained in an end-to-end manner with the aid of class-agnostic losses, frozen during the discovery phase. Further, our backbone benefits from high-level semantic features via FPN that are critical for the classification of small objects [44]. We provide more comparison details in the supplementary. We note that as our method heavily relies on the generated region proposals, we are limited by the performance of the region proposal network (i.e., we cannot detect objects not detected by the RPN network). We do not update the backbone feature extractor during self-supervised training. Successfully doing so could further boost the performance, including the localization and segmentation components.

4.4 Cross-dataset Generalization and Qualitative Results

We analyze the generalization properties of our framework by training on LVIS and evaluating on Visual Genome (VG) dataset. We present quantitative results in Table 5.

Figure 3: Visualization of predictions for validation images of the fully-supervised model and our RNCDL framework. We color the discovered novel classes in red.

Figure 4: Visualization of the common novel classes discovered in LVIS and Visual Genome datasets.

Table 5: LVIS Visual Genome comparisons with a fully-supervised method.

Method m APall m APknown m APnovel RNCDL w/ random init. 2.13 7.71 0.81 RNCDL 4.46 12.55 2.56

Fully-supervised 4.52 13.72 2.35

In Figure 3, we provide qualitative results with novel classes discovered. In Figure 4, we also visualize the commonly discovered categories. Upon manual examination of the clusters from the LVIS + VG experiment, we observe that many novel classes are synonyms, hyponyms, or hypernyms of known classes appearing in the LVIS dataset.

5 Conclusions

In this paper, we introduce an end-to-end RNCDL network for novel class discovery, detection, and localization tasks. Our model is a two-stage object detection network that can classify both instances of the labeled, known classes and those of unlabeled, novel classes. At the core of our method is a self-supervision guided by features of region proposals and a constraint for the class assignments to follow a long-tail distribution. In our experiments, we demonstrate a significant improvement over the previous state-of-the-art. Furthermore, we demonstrate the ability to detect semantic classes without any supervision at a large scale on the Visual Genome dataset.

Broader Impact

The real world contains a vast number of object categories, but labeling them all is impractical and costly. Most deep learning models for object detection are trained on datasets that cover only a fixed and limited set of categories, and they cannot generalize to novel classes that are not seen during training. This limits their applicability in scenarios where objects of interest are unknown or rare. Therefore, there is a need for methods that can discover and detect novel classes without relying on labeled data supervision. This work presents a novel end-to-end method that addresses this problem and significantly improves over previous approaches. We hope our work will inspire more research on this challenging problem and lead to more general and robust object detection models.

Acknowledgments. LLT would like to acknowledge the Humboldt Foundation through the Sofja Kovalevskaja Award. DR would like to acknowledge the CMU Argo AI Center for Autonomous Vehicle Research. We thank the anonymous reviewers for their feedback.

[1] Yuki M. Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In Neur IPS, 2020.

[2] Yuki M. Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020.

[3] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, 2018.

[4] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, 2015.

[5] Abhijit Bendale and Terrance E Boult. Towards open set deep networks. In CVPR, 2016.

[6] David M. Blei, Andrew Y. Ng, and Michael I. Jordan. Latent dirichlet allocation. JMLR, 3:993 1022, 2003.

[7] Kaidi Cao, Maria Brbi c, and Jure Leskovec. Open-world semi-supervised learning. In ICLR, 2022.

[8] Yuhang Cao, Jiaqi Wang, Ying Jin, Tong Wu, Kai Chen, Ziwei Liu, and Dahua Lin. Few-shot object detection via association and discrimination. In Neur IPS, 2021.

[9] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018.

[10] Mathilde Caron, Piotr Bojanowski, Julien Mairal, and Armand Joulin. Unsupervised pre-training of image features on non-curated data. In ICCV, 2019.

[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Neur IPS, 2020.

[12] Jianlong Chang, Lingfeng Wang, Gaofeng Meng, Shiming Xiang, and Chunhong Pan. Deep adaptive image clustering. In ICCV, 2017.

[13] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Wang, and Jia-Bin Huang. A closer look at few-shot classification. In ICML, 2019.

[14] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Neur IPS, 2013.

[15] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Image Net: A large-scale hierarchical image database. In CVPR, 2009.

[16] Akshay Raj Dhamija, Manuel Günther, and Terrance E Boult. Reducing network agnostophobia. In Neur IPS, 2018.

[17] M. Everingham, L. Van Gool, C.K.I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (VOC) challenge. IJCV, 88(2):303 338, 2010.

[18] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In CVPR, 2021.

[19] Enrico Fini, Enver Sangineto, Stéphane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In ICCV, 2021.

[20] Spyros Gidaris and Nikos Komodakis. Dynamic few-shot visual learning without forgetting. In CVPR, 2018.

[21] Ross Girshick. Fast r-cnn. In ICCV, 2015.

[22] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014.

[23] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. Technical report, California Institute of Technology, 2007.

[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019.

[25] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In ICCV, 2019.

[26] Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics. In ICLR, 2020.

[27] K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask R-CNN. In ICCV, 2017.

[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.

[29] Yen-Chang Hsu, Zhaoyang Lv, and Zsolt Kira. Deep image category discovery using a transferred similarity function. ar Xiv preprint ar Xiv:1612.01253, 2016.

[30] Yen-Chang Hsu, Zhaoyang Lv, and Zsolt Kira. Learning to cluster in order to transfer across domains and tasks. In ICLR, 2018.

[31] Ronghang Hu, Piotr Dollár, Kaiming He, Trevor Darrell, and Ross Girshick. Learning to Segment Every Thing. In CVPR, 2018.

[32] Jaedong Hwang, Seoung Wug Oh, Joon-Young Lee, and Bohyung Han. Exemplar-based open-set panoptic segmentation network. In CVPR, 2021.

[33] Lalit P Jain, Walter J Scheirer, and Terrance E Boult. Multi-class open set recognition using probability of inclusion. In ECCV, 2014.

[34] Xuhui Jia, Kai Han, Yukun Zhu, and Bradley Green. Joint representation learning and novel category discovery on single-and multi-modal data. In ICCV, 2021.

[35] K J Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, 2021.

[36] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In ICCV, 2019.

[37] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017.

[38] H. W. Kuhn and Bryn Yaw. The hungarian method for the assignment problem. Naval Res. Logist. Quart, pages 83 97, 1955.

[39] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. IJCV, 128(7):1956 1981, 2020.

[40] Suha Kwak, Minsu Cho, Ivan Laptev, Jean Ponce, and Cordelia Schmid. Unsupervised object discovery and tracking in video collections. In ICCV, 2015.

[41] Yong Jae Lee and Kristen Grauman. Object-graphs for context-aware visual category discovery. In CVPR, 2010.

[42] Yong Jae Lee and Kristen Grauman. Learning the easy things first: Self-paced visual category discovery. In CVPR, 2011.

[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014.

[44] Tsung-Yi Lin, Piotr Dollár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. In CVPR, 2017.

[45] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikäinen. Deep learning for generic object detection: A survey. IJCV, 128(2):261 318, 2020.

[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019.

[47] J. Mac Queen. Some methods for classification and analysis of multivariate observations. In Proc. Fifth Berkeley Symp. on Math. Statist. and Prob., Vol. 1, 1967.

[48] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, 2018.

[49] Aljoša Ošep, Paul Voigtlaender, Jonathon Luiten, Stefan Breuers, and Bastian Leibe. Large-scale object mining for object discovery from unlabeled video. In ICRA, 2019.

[50] Aljoša Ošep, Paul Voigtlaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In ICRA, 2020.

[51] Shafin Rahman, Salman Hameed Khan, and Fatih Porikli. Zero-shot object detection: Learning to simultaneously recognize and localize novel concepts. ACCV, 2018.

[52] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, 2017.

[53] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In Neur IPS, 2015.

[54] Michael Rubinstein, Armand Joulin, Johannes Kopf, and Ce Liu. Unsupervised joint object discovery and segmentation in internet images. In CVPR, 2013.

[55] Bryan C. Russell, William T. Freeman, Alexei A. Efros, Josef Sivic, and Andrew Zisserman. Using multiple segmentations to discover objects and their extent in image collections. In CVPR, 2006.

[56] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(7):1757 1772, 2012.

[57] Walter J Scheirer, Lalit P Jain, and Terrance E Boult. Probability models for open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11):2317 2324, 2014.

[58] Jamie Shotton, John M. Winn, Carsten Rother, and Antonio Criminisi. Texton Boost for image understanding: Multi-class object recognition and segmentation by jointly modeling texture, layout, and context. IJCV, 81(1):2 23, 2009.

[59] Josef Sivic, Bryan C. Russell, Andrew Zisserman, William T. Freeman, and Alexei A. Efros. Unsupervised discovery of visual object class hierarchies. In CVPR, 2008.

[60] Tinne Tuytelaars, Christoph H. Lampert, Matthew B. Blaschko, and Wray Buntine. Unsupervised object discovery: A comparison. IJCV, 88:284 302, 2010.

[61] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In ECCV, 2020.

[62] Huy V. Vo, Francis Bach, Minsu Cho, Kai Han, Yann Le Cun, Patrick Pérez, and Jean Ponce. Unsupervised image matching and object discovery as optimization. In CVPR, 2019.

[63] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In ECCV, 2020.

[64] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. In Neur IPS, 2021.

[65] Feng Wang, Jian Cheng, Weiyang Liu, and Haijun Liu. Additive margin softmax for face verification. IEEE Signal Processing Letters, 25(7):926 930, 2018.

[66] Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. In ICML, 2020.

[67] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. ar Xiv preprint ar Xiv:2003.06957, 2020.

[68] Zhenzhen Weng, Mehmet Giray Ogut, Shai Limonchik, and Serena Yeung. Unsupervised discovery of the long-tail in instance segmentation using hierarchical self-supervision. In CVPR, 2021.

[69] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https: //github.com/facebookresearch/detectron2, 2019.

[70] Yongqin Xian, Christoph H. Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning - a comprehensive evaluation of the good, the bad and the ugly. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018.

[71] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In ICML, 2016.

[72] Bingchen Zhao and Kai Han. Novel visual category discovery with dual ranking statistics and mutual knowledge distillation. In Neur IPS, 2021.

[73] Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, and Nicu Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021.

[74] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021.

[75] Xingyi Zhou, Vladlen Koltun, and Philipp Krähenbühl. Simple multi-dataset detection. In CVPR, 2022.