# efficient_lidar_reflectance_compression_via_scanning_serialization__14eb5ae9.pdf

Efficient Li DAR Reflectance Compression via Scanning Serialization

Jiahao Zhu * 1 Kang You * 2 Dandan Ding 1 Zhan Ma 2

Reflectance attributes in Li DAR point clouds provide essential information for downstream tasks but remain underexplored in neural compression methods. To address this, we introduce Ser Li C, a serialization-based neural compression framework to fully exploit the intrinsic characteristics of Li DAR reflectance. Ser Li C first transforms 3D Li DAR point clouds into 1D sequences via scan-order serialization, offering a device-centric perspective for reflectance analysis. Each point is then tokenized into a contextual representation comprising its sensor scanning index, radial distance, and prior reflectance, for effective dependencies exploration. For efficient sequential modeling, Mamba is incorporated with a dual parallelization scheme, enabling simultaneous autoregressive dependency capture and fast processing. Extensive experiments demonstrate that Ser Li C attains over 2 volume reduction against the original reflectance data, outperforming the state-of-the-art method by up to 22% reduction of compressed bits while using only 2% of its parameters. Moreover, a lightweight version of Ser Li C achieves 10 fps (frames per second) with just 111K parameters, which is attractive for real-world applications.

1. Introduction

Over the past two decades, the light detection and ranging (Li DAR) sensors have become indispensable components in a diverse array of applications, including autonomous driving, robotics, etc (Baur et al., 2025; Liang et al., 2024a; Li & Ibanez-Guzman, 2020). These sensors work by periodically emitting laser beams and capturing their reflections,

*Equal contribution 1School of Information Science and Technology, Hangzhou Normal University, Hangzhou, China 2School of Electronic Science and Engineering, Nanjing University, Nanjing, China. Correspondence to: Dandan Ding <Dandan Ding@hznu.edu.cn>.

Proceedings of the 42 nd International Conference on Machine Learning, Vancouver, Canada. PMLR 267, 2025. Copyright 2025 by the author(s).

0 20 40 60 80 100 120 140 160

Number of Parameters

Unicorn 111.5M

Ser Li C 2.2M

Encoding Latency (s)

G-PCC (Predlift)

Decoding Latency (s)

(a) Li DAR Point Cloud (b) Detection w/ Reflectance (c) Detection w/o Reflectance

(e) Consumption Details

Ser Li C 111K 1000x Smaller

50x Smaller

(d) Compression Performance

Ser Li C 34x

G-PCC (Predlift)

Figure 1. (a-c) Reflectance plays an indispensable role in downstream tasks such as 3D object detection; (d) The proposed Ser Li C establishes a new state-of-the-art by outperforming the latest compression standard G-PCC (Zhang et al., 2024b) and the learningbased work Unicorn on widely accepted datasets KITTI (Behley et al., 2019), Ford (Pandey et al., 2011), and nu Scenes (Caesar et al., 2020) (the number denotes sequence number); (e) Ser Li C achieves ultra-low coding latency with an exceptionally lightweight model.

producing vast amounts of data points that form the point cloud a digital representation of the surrounding 3D environment for downstream perception tasks (Chen et al., 2024; Hu et al., 2023). The massive volume of such data, often reaching gigabytes per minute, highlights the development of efficient Li DAR point cloud compression (PCC) techniques to enable effective storage and transmission (Li et al., 2024; Zhang et al., 2024b; Liu et al., 2020).

Point cloud reflectance, derived from the intensity of returned laser pulses in Li DAR, serves as a crucial descriptor of the physical properties of objects. It is widely used in downstream tasks such as object detection and semantic segmentation (Viswanath et al., 2024; Tatoglu & Pochiraju, 2012). For instance, our experiments show that removing the reflectance causes a dramatic drop in pedestrian (and cyclist) detection accuracy incorporating the pre-trained Point Pillar (Lang et al., 2019) detection model, having AP (average precision) from 51.4 (62.8) to 14.1 (34.3) on the KITTI dataset (see Figure 1), making it impractical for use.

Efficient Li DAR Reflectance Compression via Scanning Serialization

While existing neural compression methods have achieved significant performance gains, they mostly focus on the geometry or RGB color attributes (Wang et al., 2025; You et al., 2024; Zhang et al., 2023; Wang et al., 2023), and the compression of Li DAR reflectance is relatively underexplored with uncompetitive efficiency. This compression gap is from the intrinsic challenges of reflectance data.

The first challenge lies in the inherent sparsity of Li DAR scans. Unlike object point clouds comprising relatively dense points, Li DAR point clouds are much sparser, as laser pulses are reflected only from a limited subset of distant points in the environment (Kong et al., 2023; Raj et al., 2020). To address this, convolution-based methods (Wang et al., 2025; 2023) have to employ large-kernel convolutions (e.g., 73) to increase the receptive field to aggregate sufficient points in 3D space for analysis. However, this incurs an expensive computational burden while yielding only marginal performance gains, as the fixed-size large kernel remains insufficient to capture dynamic context.

The second challenge is the intricate physics of underlying Li DAR reflectance. Reflectance intensity is influenced by various factors, such as the material properties of objects, the angle of incidence of laser pulses, the distance between the sensor and the target, and hardware variations (Liang et al., 2024a; Viswanath et al., 2024; Fang et al., 2014). These introduce intertwined correlations across points that extend beyond simple spatial proximity. For instance, reflectance values from surfaces with similar material properties may exhibit strong correlations even when spatially distant. However, existing methods largely neglect these physical interactions inherent in Li DAR reflectance. They typically adopt modeling strategies designed for color attributes in object point clouds and rely on geometric 3D spatial proximity to define Li DAR point relationships. Consequently, their performance remains limited.

Therefore, this work proposes a serialization-based neural compression framework, dubbed Ser Li C, that utilizes sequence modeling to improve the efficiency of Li DAR reflectance compression. In contrast to current techniques that directly work in 3D space for spatial correlation exploration, Ser Li C follows the Li DAR scan order to serialize 3D points into 1D point sequences, for which expensive 3D operations are removed. The serialization aligns with the Li DAR laser scanning mechanism, allowing for a device-centric rather than spatial perspective in reflectance analysis. Upon this representation, we devise an entropy model to capture point dependencies in a window of each sequence for effective contextual modeling. Within each window, the selective state space model (a.k.a. Mamba (Gu & Dao, 2023)) is employed for autoregressive coding. Additionally, Li DAR scanning-relevant information is derived based on the input point cloud as context to improve the modeling accuracy.

Extensive experiments on benchmark datasets show that Ser Li C achieves remarkably high performance, over 2 volume reduction against the original reflectance data and up to 22% bit rate reduction compared to the state-of-the-art Unicorn method (Wang et al., 2025), while using only 2% of the parameters and 10% of the GPU memory, as illustrated in Figure 1. Moreover, Ser Li C is hardware-friendly, relying solely on simple networks rather than specific convolution libraries (e.g., Minkowski Engine (Choy et al., 2019)) commonly used in existing methods. Our key contributions are summarized as follows:

We propose Ser Li C, a lossless reflectance compression method for Li DAR point clouds, leveraging scan-order serialization to transform a 3D point cloud to 1D point sequences for efficient representation.

We generate Li DAR information (scanning index and radial distance) for each point, along with the previous decoded reflectance, as context to exploit point dependencies in a sequence, supported by the selective state space model with a dual parallelization mechanism.

Ser Li C delivers notable performance on benchmark datasets, offering high compression efficiency, ultralow complexity, and strong robustness. Its light version runs 30 fps with frame pipelining and 10 fps without, with only 111K model parameters.

2. Related Work

2.1. Serialization-based Point Cloud Analysis

Transformers. Serialization has shown remarkable efficacy in point cloud analysis tasks (Wu et al., 2024; Liu et al., 2023; Wang, 2023), owing to the inherent simplicity and computational efficiency of structured data representations. Notable studies, including Oct Former (Wang, 2023), Flat Former (Liu et al., 2023), and Point Transformer v3 (Wu et al., 2024), have utilized serialization-based Transformer architectures to implement attention mechanisms in structured spaces, achieving high-performance representation. However, Transformers bring quadratic computational complexity, which poses challenges in long sequence modeling (Liu et al., 2024).

Mamba. Recently, Mamba (Gu & Dao, 2023) introduced the selective state space model (SSM) mechanism, garnering significant attention for its linear computational complexity and superior performance. Numerous works (Liang et al., 2024b; Zhang et al., 2025; Han et al., 2024; Zhang et al., 2024a; Wang et al., 2024) have successfully integrated Mamba into point cloud analysis to enhance efficiency. For instance, Point Mamba (Liang et al., 2024b) applies selective SSM in point cloud analysis by reorganizing key point features into sequences for efficient processing within the

Efficient Li DAR Reflectance Compression via Scanning Serialization

Mamba framework. Extending this paradigm, several studies (Zhang et al., 2024a; Zeng et al., 2024) have introduced Mamba to Li DAR point clouds. However, their serialization approaches are limited to traditional space-filling curves (Morton, 1966; Hilbert, 2013), overlooking unique scanning mechanisms and physical information inherent in Li DAR point clouds.

2.2. Point Cloud Attribute Compression

Rules-based Methods. Recent advancements in Point Cloud Attribute Compression (PCAC) have led to the standardization of techniques, with MPEG s G-PCC (Zhang et al., 2024b) serving as the benchmark. G-PCC incorporates two primary methods for Li DAR reflectance compression: Region-Adaptive Hierarchical Transform (RAHT) (De Queiroz & Chou, 2016) and Predicting/Lifting Transform (Predlift) (Mammou et al., 2018). To address GPCC s latency and complexity, L3C2 (S ebastien & Jonathan, 2021) was introduced in 2021, specifically tailored for Li DAR point clouds. However, its dependence on detailed sensor configurations (e.g., the precise pitch angles of individual lasers) limits its applicability.

Learning-based Methods. Neural models have revolutionized PCAC by incorporating deep-learning techniques to capture attribute dependencies. CNe T (Nguyen & Kaup, 2023) employs an autoregressive framework with causal priors for sequential prediction, achieving high compression gains at the expense of extremely high computational cost. Building on this, MNe T (Nguyen et al., 2023) adopts a multiscale strategy to enhance computational efficiency but the compression gain is largely compromised. Additionally, CNe T and MNe T are deigned for color attributes, failing to compress the Li DAR reflectance. Po Lo PCAC (You et al., 2024) introduces a point-based compression pipeline that models attributes in groups based on antecedents; however, it shows certain dependence on training data, limiting its generalizability. Unicorn (Wang et al., 2025), the latest framework, introduces a universal multiscale coding mechanism using 3D sparse convolution library Minkowski Engine (Choy et al., 2019) to predict attribute values across multiple scales, achieving state-of-the-art performance. To attain high performance, Unicorn devise large kernel sizes in 3D convolution, resulting in intensive complexity.

Remarks. While these learning-based methods have demonstrated promising performance improvements, they mainly focus on general-purpose attributes (e.g., color intensities), which limits their effectiveness in specialized scenarios such as Li DAR reflectance. As the industrial adoption of Li DAR technology grows across various applications like autonomous driving and urban planning (Liang et al., 2024a), developing specialized codecs for Li DAR point clouds is essential to meet practical requirements.

3.1. Problem Definition

Given a Li DAR point cloud with N points, we denote its reflectance intensities as X = {xi}N i=1 and the geometry coordinates as C = {ci}N i=1. Here, xi and ci represent the reflectance attribute and the spatial coordinate of the i-th point. Existing approaches (Wang et al., 2025; 2023; You et al., 2024) rely exclusively on the unordered geometric coordinates C to model point correlations, thereby neglecting the inherent sequential scanning characteristics of Li DAR data. In contrast, this paper introduces serialization for leveraging the sequential nature of Li DAR scans. This is because 1) Serialization transforms point clouds into a 1D representation, enabling more efficient and structured modeling from a device-centric perspective; and 2) Serialization supports the efficient processing of information that more closely reflects the physical characteristics of Li DAR scanning (C C ), resulting in more accurate analysis.

The serialization process reorganizes the input point cloud (X, C) into L point sequences as follows (see Figure 2):

{(Xl, C l )}L l=1 = Serialize (X, C) , (1)

where L denotes the number of laser beams emitted by the Li DAR sensor. Each sequence Xl = {xl,i}Nl i=1 represents the reflectance intensities from the l-th laser beam, and

C l = n c l,i o Nl

i=1 encodes the contextual information of the corresponding points.

In this context, we define Li DAR information c l,i = (vl,i, ul,i, ρl,i), including the laser index vl,i, the azimuth angle index ul,i, and the radial distance ρl,i for each point. These information can be entirely derived from the geometric coordinates C during the serialization process, as detailed in Section 3.2.

With the serialized representation (Xl, C l ), our objective is to design a parameterized entropy model that incorporates the priors C l to exploit spatial and contextual dependencies in the reflectance data Xl. This model aims to approximate the conditional probability distribution p(Xl|C l ) as closely as possible to the true distribution q(Xl|C l ). The total encoded bit rate R is the sum of bit rates for all sequences, which can be expressed using the Shannon cross-entropy between two distributions:

l=1 EXl q(Xl|C l )[ log2 p(Xl|C l )]. (2)

3.2. Serialization

Serialization has demonstrated significant efficacy in a wide range of point cloud analysis tasks, owing to the inherent simplicity and computational efficiency of structured

Efficient Li DAR Reflectance Compression via Scanning Serialization

Li DAR Point Cloud Scan-order Serialization

Point Sequences

Mamba-driven Autoregressive Coding

Mamba Block Mamba Block Mamba Block

Mamba Block Mamba Block Mamba Block

Enc. & Dec. Process

Enc. Process

Dec. Process

AE Arithmetic Encoder

AD Arithmetic Decoder

Window-based Parallelization

Parallelized

Parallelized

Figure 2. Ser Li C Framework. The input 3D Li DAR point cloud is first serialized into 1D ordered point sequences, which are then divided into windows for parallel processing. For each window, a Mamba-driven autoregressive coding (MDAC) scheme is employed, which embeds scanning index (Fpos i ), radial distance (Fρ i ), and prior reflectance (Fx i 1) as context to generate the probability mass function (PMF) for the reflectance intensity of the target (i-th) point.

data (Liang et al., 2024b; Zhang et al., 2025; Wu et al., 2024). Unlike conventional point cloud serialization strategies, such as the space-filling curves (Morton, 1966; Hilbert, 2013; Liang et al., 2024b; Zhang et al., 2025), our proposed Ser Li C introduces a Li DAR scan-order serialization, specifically designed to effectively preserve and leverage the intrinsic regularities within a Li DAR point cloud.

Coordinate Mapping. Specifically, the Cartesian coordinate ci of a point is first converted to spherical coordinate space, yielding the elevation angle φi π

2 , azimuth angle ϕi ( π, π], and radial distance ρi (0, + )1. The transformation is expressed as:

x2 i + y2 i + z2 i

φi = arcsin zi

ϕi = atan2 (yi, xi)

where xi, yi, and zi denote the Cartesian coordinate of ci. Next, each point is mapped to a discrete grid by assigning a laser index vi and an azimuthal index ui, respectively, based on its elevation angle φi and azimuth angle ϕi:

vi = L φi φdown

1In practical systems, the maximum detection range of Li DAR is constrained by sensor specifications, typically up to 400 meters (Liang et al., 2024a).

where L and W represent the angular resolutions of the elevation and azimuth angles, respectively; L is the number of laser channels; denotes the floor operation; φup and φdown are the maximum and minimum elevation angles, defining the upward and downward view field of Li DAR sensor. Notably, the values of vi and ui are clipped to ranges [1, L] and [1, W], respectively. As such, the contextual information c i of each point is aggregated as (vi, ui, ρi).

Reordering. Based on the angular indices vi and ui, the point cloud data are organized and sorted as follows:

Points with the same laser index vi = l(l [1, L]) are aggregated into subsets ( Xl, C l ), with ( Xl, C l ) = {(xi, c i ) | vi = l, i [1, N]}.

Within each subset, points are ordered by increasing azimuthal index ui, yielding a sequence (Xl, C l ) = {(xi, c i ) ( Xl, C l ) | ui uj, i < j}.

In this way, the proposed serialization method restructures the point cloud into internally ordered sequences, enabling efficient autoregressive coding using the state-space model.

3.3. Contextual Construction

Figure 2 illustrates the state space model-based autoregressive encoder for sequentially compressing the reflectance attributes of the point sequence. Each point is tokenized into a contextual combination of scanning index, radial distance, and prior reflectance, with the aim of exploring correlations to its neighbor for accurate probability estimation.

Efficient Li DAR Reflectance Compression via Scanning Serialization

Scanning Index. Previous methods usually explicitly leverage geometry coordinates to identify correlated neighbors for a point. However, spatially close points may not be highly correlated as the Li DAR points are derived following its unique scanning mechanism. To this end, leveraging the scanning index for point context will yield better results since points with the same index are captured similarly. Specifically, the scanning index feature Fpos i of the i-th point is derived by directly embedding the laser and azimuth indices (i.e., vi and ui) obtained in Section 3.2:

Fpos i = Embed (vi) Embed (ui) , 1 i Nl, (5)

where Embed denotes the embedding layer, and refers to the concatenate operation.

Radial Distance. Li DAR reflectance essentially represents the backscattered intensity of Li DAR signals, which are intrinsically influenced by the distance between the object and the sensor, based on the light transmission theory (Viswanath et al., 2024; Fang et al., 2014). Therefore, Ser Li C explicitly incorporates radial distance as context to enhance the reflectance compression. Specifically, the radial distance ρi (see Eq. (3)) is normalized and transformed into a high-dimensional spatial feature Fρ i using a Linear layer:

Fρ i = Linear (Norm (ρi)) , 1 i Nl, (6)

where Norm refers to the min-max normalization that rescales ρi to the range of (0, 1).

Prior Reflectance. Naturally, the reflectance value of previous point in the sequence is embedded (Fx i 1) as context.

We then combine all context for subsequent processing:

Ftoken i = Fpos i Fρ i Fx i 1 , 1 i Nl. (7)

For the case of i=1, where prior reflectance is unavailable, the value of Fx 0 is estimated by the neural network.

3.4. Entropy Coding

Mamba Coder. Given the collected context token Ftoken i of a point sequence, Mamba is ideally suited for autoregressive coding due to its ability to model sequential dependencies and capture correlations between points.

Thus, the token Ftoken i is passed through several Mamba blocks. A standard Mamba layer is shown in Figure 3(a), and the processing flow can be summarized as follows:

Fs i = Layer Norm Fs 1 i , Fs i = σ DWConv Linear Fs i , ˆFs i = σ Linear Fs i ,

Fs i = Linear Selective SSM( Fs i) ˆFs i + Fs 1 i ,

Mamba Block N

(a) Mamba Block

Attention Block N

Masked Multi Head Attention

Feed-forward

(b) Attention Block

Figure 3. Basic Mamba block and Attention block. LN refers to Layer Norm; σ denotes Si LU activation; means Hadamard product; represents element-wise addition.

where Fs i denotes the output of the s-th Mamba layer; DWConv represents the depth-wise convolution; σ refers to the Si LU activation function; means Hadamard product. The initial input F0 i is set to the token feature Ftoken i .

Let Fout i be the output of the final Mamba layer. Then a classifier is used to predict the probability mass function (PMF) for the target (i-th) point, leveraging the Linear transform and Softmax function:

pl,i | xl,<i, c l, i = Softmax Linear Fout i . (9)

Dual Parallelization. Ser Li C proposes a dual parallelization strategy at both the sequence and window levels to accelerate the coding process. First, as shown in Figure 2, Ser Li C processes point sequences in parallel, disregarding correlations across sequences. This is justified, as point sequences are scanned by separate laser sensor rotations, inherently lacking strong inter-sequence dependencies. As a result, sequence parallelization does not compromise coding efficiency but greatly enhances processing speed.

Furthermore, Ser Li C introduces window-based parallelization within each point sequence. Li DAR scanning systems, particularly those utilizing mechanical rotation, produce sequences with extensive spatial coverage, often spanning a full 360-degree field of view. Within such sequences, spatial regions exhibit varying degrees of correlation local regions like the area directly in front of the sensor have stronger internal correlations than distant regions located behind or to the sides. This localized correlation characteristic enables

Efficient Li DAR Reflectance Compression via Scanning Serialization

Table 1. Quantitative compression gains against other methods

BITS PER POINT (BPP) COMPRESSION GAIN G-PCC G-PCC UNICORN SERLIC SERLIC VS. VS. VS. VS. (RAHT) (PREDLIFT) (LIGHT) RAHT PREDLIFT UNICORN LIGHT

11 5.10 5.01 4.59 4.06 3.80 -25.49% -24.15% -17.21% -6.40% 12 4.56 4.51 4.14 3.68 3.45 -24.34% -23.50% -16.67% -6.25% 13 4.79 4.62 4.15 3.63 3.35 -30.06% -27.49% -19.28% -7.71% 14 5.24 5.22 4.82 4.35 4.16 -20.61% -20.31% -13.69% -4.37% 15 5.11 5.03 4.61 4.13 3.87 -24.27% -23.06% -16.05% -6.30% 16 5.00 4.96 4.55 4.11 3.85 -23.00% -22.38% -15.38% -6.33% 17 4.80 4.73 4.38 3.87 3.64 -24.17% -23.04% -16.89% -5.94% 18 5.06 4.97 4.56 4.08 3.80 -24.90% -23.54% -16.67% -6.86% 19 4.63 4.53 4.11 3.50 3.20 -30.89% -29.36% -22.14% -8.57% 20 4.51 4.35 3.98 3.52 3.28 -27.27% -24.60% -17.59% -6.82% 21 4.83 4.78 4.41 3.93 3.69 -23.60% -22.80% -16.33% -6.11% AVG. 4.88 4.79 4.39 3.90 3.64 -25.41% -24.01% -17.08% -6.67%

02 5.18 5.03 4.89 4.29 3.57 -31.08% -28.94% -26.99% -16.78% 03 5.13 5.05 5.04 4.46 4.11 -19.91% -18.65% -18.45% -7.85% AVG. 5.16 5.04 4.97 4.29 3.84 -25.58% -23.81% -22.74% -12.33%

01 3.93 3.61 - 3.13 2.89 -26.46% -19.94% - -4.30% 02 3.41 3.08 - 2.46 2.58 -33.14% -25.97% - -7.32% 03 3.18 2.77 - 2.23 2.34 -34.59% -24.91% - -6.73% 04 3.05 2.69 - 2.28 2.41 -22.62% -12.27% - -3.51% 05 4.40 4.04 - 3.23 3.33 -31.82% -25.74% - -7.12% AVG. 3.59 3.24 - 2.78 2.52 -29.81% -22.22% - -9.35%

Table 2. Quantitative compression gains against L3C2

CLASS SEQ. BPP CR GAIN L3C2 SERLIC VS. L3C2

02 4.83 3.57 -26.15% 03 4.98 4.11 -17.39% AVG. 4.90 3.84 -21.63%

effective compression even when processing is limited to smaller windows rather than the entire sequence. Accordingly, Ser Li C partitions each sequence into independent windows for parallel processing, ensuring both efficiency and performance. Figure 2 depicts the window slicing method. Each sequence is partitioned equally. Padding points are appended at the end to maintain a uniform window size for consistent processing.

4. Experiment and Analysis

4.1. Datasets

We conducted experiments on well-known Li DAR datasets, including KITTI (Behley et al., 2019), Ford (Pandey et al., 2011), and nu Scenes (Caesar et al., 2020).

KITTI or Semantic KITTI is a large-scale Li DAR dataset used for semantic scene understanding. It contains 22 sequences, a total of 43,552 frames of outdoor scenes collected using the Velodyne HDL-64E Li DAR

sensor. There are around 120k points on average per frame. These raw floating-point coordinates are quantized to 1mm precision (18 bits) with 7-bit reflectance.

Ford is also collected using Velodyne HDL-64E. The common test condition (CTC) defined by MPEG (WG 07 MPEG 3D Graphics Coding and Haptics Coding, 2024b) utilizes three Ford sequences, each having 1,500 frames at 1mm precision with 8-bit reflectance. The first sequence is for training, and the remaining two are for testing.

nu Scenes is a large-scale dataset collected for autonomous driving using Velodyne HDL-32E. It has ten subsets, each containing 85 scenes. For training, we extract the first 100 frames from the first 12 scenes in the first five subsets, resulting in 6,000 frames. For testing, we select the first 90 frames from the first scene of each of the last five subsets, yielding 450 frames numbered as sequences #01 to #05. Also, we quantize them to 1mm geometric precision with 8-bit reflectance.

4.2. Experimental Details

Training Setting. We implement Ser Li C using Python 3.10 and Py Torch 2.5. The model is trained with Adam W optimizer (Loshchilov & Hutter, 2019), using a learning rate of 2 10 4 and a batch size of 64. We employ a cosine annealing strategy (Loshchilov & Hutter, 2017) to gradually reduce the learning rate to 5 10 5. The model is randomly initialized and trained for 25 epochs for each dataset. For

Efficient Li DAR Reflectance Compression via Scanning Serialization

fair comparisons, all experiments are conducted on the same platform, equipped with an NVIDIA RTX 4090 GPU, an Intel Core i9-13900K CPU, and 64GB of memory.

The optimization objective is to minimize the bit rate (as defined in Equation (2)) for transmitting reflectance data, which can be further formulated as the negative loglikelihood of the observed reflectance values across all sequences l [1, L] and points i [1, Nl]:

i=1 log2 pl,i (xl,i | contextl,i) , (10)

where pl,i (xl,i | contextl,i) denotes the conditional probability estimated by the parametrized entropy model pl,i based on the contextl,i.

Testing Setting. The testing conditions strictly follow the CTC of MPEG AI-PCC (WG 07 MPEG 3D Graphics Coding and Haptics Coding, 2024a). Quantitative analysis is measured in bits per point (bpp) and compression ratio (CR). The latest G-PCC version TMC13v232, which provides state-of-the-art performance through RAHT and Predlift modes, is compared. Also, we compare Ser Li C with L3C2 (S ebastien & Jonathan, 2021), a profile dedicated for Li DAR PCC in MPEG, and existing learning-based Unicorn (Wang et al., 2025) which shows superior performance on Li DAR reflectance compression. All methods are evaluated under the same training/testing datasets and conditions.

4.3. Compression Performance

Table 1 presents a detailed comparison of the overall bit rate and CR gains of Ser Li C against G-PCC (RAHT), G-PCC (Predlift), and Unicorn. As observed, Ser Li C consistently outperforms both G-PCC (RAHT) and G-PCC (Predlift), across all three tested datasets, e.g., Ser Li C rivals G-PCC (RAHT) by 25.41%, 25.58%, and 29.81% on three datasets. Substantial performance improvements (17.08% on KITTI and 22.74% on Ford) are observed when compared with Unicorn and L3C2 (21.63% on Ford).

Notice that the compression of L3C2 relies heavily on detailed parameters of Li DAR sensor, including the elevation angle of each laser sensor and their respective distances to the Li DAR center. For datasets lacking such sensor data, such as KITTI and nu Scenes, L3C2 cannot function effectively (hence we only provide L3C2 results on Ford). In contrast, Ser Li C only requires basic Li DAR information such as the angular resolutions of the elevation and azimuth angles. These results are evidence of the superior effectiveness of Ser Li C as well as its robustness on various datasets.

2https://github.com/MPEGGroup/ mpeg-pcc-tmc13

Table 3. Computational complexity on the KITTI dataset

METHOD MEM. PARAM. ENC. DEC.

RAHT - - 0.36S 0.34S PREDLIFT - - 0.74S 0.74S

UNICORN 4.36GB 111.5M 2.77S 2.77S SERLIC 0.41GB 2.2M 0.09S /0.18S 0.13S /0.23S SERLIC (LIGHT) 0.25GB 111K 0.03S /0.08S 0.03S /0.09S

* RESULTS OF USING THE FRAME-LEVEL PIPELINING

4.4. Computational Complexity

Table 3 reports the computation complexity. Ser Li C uses only 2.2M parameters, which is only 2% of Unicorn.

For runtime, the coding process consists of three components: context construction (CC, on CPU), neural network (NN, on GPU), and arithmetic coding (AC, on CPU). By utilizing a three-stage pipeline in frame-level to arrange these components, Ser Li C achieves an encoding/decoding speed of 0.09/0.13 seconds per frame, i.e., >11/7 frames per second (fps). Even when the frame-level pipelining is disabled, the total runtime remains much shorter (<10%) than that of Unicorn, achieving 4 to 5 fps. Note that our experiments use point clouds with 18-bit geometry, containing hundreds of thousands of points per frame. If applied to 12-bit point clouds, which have far fewer points but still maintain high accuracy for downstream tasks, the coding speed of Ser Li C would be even faster. Detailed runtime and analysis can be found in our supplementary material.

Regarding GPU memory usage, Ser Li C uses 0.41 GB while Unicorn uses 4.36 GB, only 9% of Unicorn. The number of parameters in Ser Li C is also much smaller: the full version is only 2% of Unicorn while the light version is only 1 . All these confirm the ultra-low complexity of Ser Li C.

4.5. Ablation Study

Ablation studies are conducted on the KITTI dataset to validate Ser Li C. Default settings are marked in gray . The results of CR gains are over the G-PCC (RAHT) anchor.

Table 4. Ablation study on contextual construction

VARIANTS BPP CR GAIN + Fx i 1 + Fρ i + Fpos i 3.92 -19.67% 3.82 -21.72% 3.73 -23.57% 3.64 -25.41%

Contextual Construction. When certain components are disabled, we expand the embedding dimension of the remaining components to ensure that the overall network di-

Efficient Li DAR Reflectance Compression via Scanning Serialization

mension entering the Mamba block remains consistent. In Table 4, when both radial distance and scanning index embeddings are disabled, the only available prior information is the reflectance of previous point Fx i 1. As a result, the gains decrease from 25.41% to 19.67%. Also, disabling either Fρ i or Fpos i causes loss. These results confirm the significance of utilizing Li DAR physical information.

Table 5. Ablation study on parallel window size, maximum GPU memory (Mem.), neural network (NN) latency (encoding/decoding), and total coding time (encoding/decoding)

SIZE BPP MEM. (GB) NN (S) TOTAL (S)

64 3.68 0.61 0.08/0.09 0.16/0.18 128 3.64 0.41 0.09/0.13 0.18/0.23 256 3.64 0.32 0.17/0.28 0.32/0.43 512 3.64 0.29 0.33/0.57 0.57/0.60 1024 3.66 0.28 0.66/1.03 1.07/1.43

Window-based Parallelization. We further investigate the impact of window-based parallelization on encoding/decoding runtime and memory usage. For fair comparisons, we disable the frame pipelining in this ablation. Table 5 presents the results for various parallel window sizes, ranging from 64 to 1024. Increasing the window size substantially extends the encoding and decoding time due to the involvement of more autoregressive steps in a window. At the same time, GPU memory usage decreases as the number of parallel windows is reduced. As a result, we set the window size as 128 by default for Ser Li C, striking a balance between coding time and memory consumption.

Table 6. Ablation study on the Mamba network

MODULE SETS. PARAM. BPP CR GAIN

1 488K 3.84 -21.31% 3 1.4M 3.68 -24.48% 5 2.2M 3.64 -25.41% 7 3.1M 3.61 -26.02%

64 176K 3.76 -22.95% 128 609K 3.69 -24.39% 256 2.2M 3.64 -25.41% 512 8.6M 3.61 -26.02%

Mamba Network. To study the effect of the number of Mamba layers on model performance, we conducted experiments with layer configurations {1, 3, 5, 7}. For each configuration, the network dimension is set to 256. As shown in Table 6, the CR gains against G-PCC (RAHT) increase from 21.31% to 26.02% as the number of layers increases from 1 to 7. We also investigate the impact of network dimension by setting it to {64, 128, 256, 512}. Here, the number of Mamba layers is fixed at 5. As reported in Table 6, the gains increase from 22.95% to 26.02% as the network dimension

64 128 256 512 1024

Window Size

Decoding Latency (s)

(a) Decoding Latency

64 128 256 512 1024

Window Size

GPU Memory (GB)

(b) Maximum GPU Memory

Figure 4. Comparison of decoding latency and running memory between Attention and Mamba implementations in Ser Li C.

increases from 64 to 512. These results are in line with the general expectation that larger models have better capacity. However, higher complexity is required.

Attention-based Coding. Attention represents another prevalent neural structure for sequence modeling, as exemplified by Oct Attention (Fu et al., 2022) and Point Transformer v3 (Wu et al., 2024). For a fair comparison, we substitute the original Mamba module with Masked Multi-Head Attention (Vaswani et al., 2017) (detailed in Figure 3(b)) while maintaining strict consistency in key hyperparameters, such as the number of layers and channels. We observe that using Attention receives similar coding gains to Mamba but higher complexity. Figure 4 illustrates the complexity of Attention-based and Mamba-based Ser Li C under different configurations. It is observed that Attention has much higher complexity, in both decoding delay and GPU memory consumption. This is because Attention computes point dependencies within a sequence window W via matrix multiplication, resulting in quadratic complexity Θ(W 2). Particularly for decoding which requires point-by-point processing in a window, the complexity increases to Θ(W 3).

In contrast, Mamba consistently maintains linear computational complexity. For example, when the parallel window size is 128, Attention needs 5.12 seconds and 1.46 GB memory for decoding a KITTI point cloud, while Mamba needs only 0.23 seconds with 0.41 GB memory, which is much more beneficial for resource-constrained applications.

Light Version. A light version of Ser Li C is implemented by reducing the network dimension of standard Ser Li C from 256 to 64, the number of layers from 5 to 3, and window size from 128 to 32. As presented in Table 1, our Ser Li C (light) receives 6.67% loss against the standard version on the KITTI dataset. However, the computational complexity is greatly reduced, with the number of parameters substantially decreased to just 111K over 1,000 smaller than that of Unicorn and the coding speed at > 10 fps (0.08-0.09 seconds per frame) even without the frame pipelining implementation. Under the frame-level pipeline, Ser Li C (light) achieves a real-time processing speed of over 30 fps (0.03 seconds per frame).

Efficient Li DAR Reflectance Compression via Scanning Serialization

Table 7. Quantitative compression gains against G-PCC (RAHT) on the non-rotational dataset

CLASS SEQ. BPP CR GAIN RAHT SERLIC VS. RAHT

02 3.71 2.28 -38.54% 03 3.82 2.25 -41.10% AVG. 3.77 2.27 -39.71%

Table 8. Quantitative compression gains against G-PCC (RAHT) on various density scenarios in KITTI

CLASS SEQ. BPP CR GAIN RAHT SERLIC VS. RAHT

CITY 4.67 3.24 -30.62% RESIDENTIAL 4.96 3.81 -23.19% ROAD 5.41 4.39 -18.85% CAMPUS 4.86 3.43 -29.42% AVG. 4.98 3.72 -25.28%

Non-Rotational Adaptation. In addition to the KITTI, Ford, and nu Scenes datasets that are captured by rotational Li DAR scanning, we also evaluated Ser Li C on the nonrotational Innoviz QC dataset provided by MPEG. It provides three sequences, each having 300 frames with 16-bit coordinates and 8-bit reflectance. We used the first sequence for training and the remaining two for testing. Results shown in Table 7 demonstrate Ser Li C s strong generalization capability to non-rotational Li DAR systems.

Scenario Robustness. To further validate the robustness of Ser Li C, we provide its detailed performance on KITTI across various scenarios, as reported in Table 8. Results indicate that Ser Li C consistently outperforms G-PCC in all scenes. Relatively, Ser Li C achieves better performance in City and Campus scenes compared to Residential and Road scenes. This occurs due to dense roadside vegetation interfering with sensor-recorded reflectance characteristics and thus increasing compression difficulty.

5. Conclusion

This paper presents Ser Li C, a serialization-based neural compression framework tailored for Li DAR reflectance attribute. By leveraging scan-order serialization, Ser Li C transforms 3D point clouds into 1D sequences, aligning with Li DAR scanning mechanisms and enabling efficient sequential modeling. The Mamba model with physics-informed tokenization further enhances its ability to capture points correlations autoregressively, while maintaining linear-time complexity. Its high efficiency, ultra-low complexity, and strong robustness make it a practical solution for real Li DAR applications. Future work will extend Ser Li C to lossy compression for higher compression efficiency.

Impact Statement

This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here.

Acknowledgments

This research was supported by National Natural Science Foundation of China (62171174) and Natural Science Foundation of Jiangsu Province (BK20243038). We are grateful to the anonymous reviewers for comments on early drafts of this paper. We also extend our sincere thanks to the authors of the relevant works used in our comparative studies for providing the latest results for evaluation.

Baur, S. A., Moosmann, F., and Geiger, A. LISO: Li DARonly self-supervised 3D object detection. In European Conference on Computer Vision, pp. 253 270. Springer, 2025.

Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., and Gall, J. Semantic KITTI: A dataset for semantic scene understanding of Li DAR sequences. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 9296 9306, 2019.

Caesar, H., Bankiti, V., Lang, A. H., Vora, S., Liong, V. E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., and Beijbom, O. nu Scenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11621 11631, 2020.

Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., and Li, H. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.

Choy, C., Gwak, J., and Savarese, S. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3075 3084, 2019.

De Queiroz, R. L. and Chou, P. A. Compression of 3D point clouds using a region-adaptive hierarchical transform. IEEE Transactions on Image Processing, 25(8):3947 3956, 2016.

Fang, W., Huang, X., Zhang, F., and Li, D. Intensity correction of terrestrial laser scanning data by estimating laser transmission function. IEEE Transactions on Geoscience and Remote Sensing, 53(2):942 951, 2014.

Efficient Li DAR Reflectance Compression via Scanning Serialization

Fu, C., Li, G., Song, R., Gao, W., and Liu, S. Oct Attention: Octree-based large-scale contexts model for point cloud compression. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pp. 625 633, 2022.

Gu, A. and Dao, T. Mamba: Linear-time sequence modeling with selective state spaces. ar Xiv preprint ar Xiv:2312.00752, 2023.

Han, X., Tang, Y., Wang, Z., and Li, X. Mamba3D: Enhancing local features for 3D point cloud analysis via state space model. In Proceedings of the 32nd ACM International Conference on Multimedia, pp. 4995 5004, 2024.

Hilbert, D. Dritter Band: Analysis Grundlagen der Mathematik Physik Verschiedenes: Nebst Einer Lebensgeschichte. Springer-Verlag, 2013.

Hu, S., Liu, W., Li, M., Zhang, Y., Liu, X., Wang, X., Zhang, L. Y., and Hou, J. Point CRT: Detecting backdoor in 3D point cloud via corruption robustness. In Proceedings of the 31st ACM International Conference on Multimedia, pp. 666 675, 2023.

Kong, L., Liu, Y., Chen, R., Ma, Y., Zhu, X., Li, Y., Hou, Y., Qiao, Y., and Liu, Z. Rethinking range view representation for Li DAR segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 228 240, 2023.

Lang, A. H., Vora, S., Caesar, H., Zhou, L., Yang, J., and Beijbom, O. Point Pillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 12697 12705, 2019.

Li, G., Gao, W., and Gao, W. Point Cloud Compression: Technologies and Standardization. Springer Nature, 2024.

Li, Y. and Ibanez-Guzman, J. Li DAR for autonomous driving: The principles, challenges, and trends for automotive Li DAR and perception systems. IEEE Signal Processing Magazine, 37(4):50 61, 2020.

Liang, D., Zhang, C., Zhang, P., Liu, S., Li, H., Niu, S., Rao, R. Z., Zhao, L., Chen, X., Li, H., et al. Evolution of laser technology for automotive Li DAR, an industrial viewpoint. nature communications, 15(1):7660, 2024a.

Liang, D., Zhou, X., Xu, W., Zhu, X., Zou, Z., Ye, X., Tan, X., and Bai, X. Point Mamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024b.

Liu, H., Yuan, H., Liu, Q., Hou, J., and Liu, J. A comprehensive study and comparison of core technologies for

MPEG 3D point cloud compression. IEEE Transactions on Broadcasting, 66(3):701 717, 2020.

Liu, X., Zhang, C., and Zhang, L. Vision Mamba: A comprehensive survey and taxonomy. ar Xiv preprint ar Xiv:2405.04404, 2024.

Liu, Z., Yang, X., Tang, H., Yang, S., and Han, S. Flat Former: Flattened window attention for efficient point cloud transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1200 1211, 2023.

Loshchilov, I. and Hutter, F. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, 2017.

Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, 2019.

Mammou, K., Tourapis, A., Kim, J., Robinet, F., Valentin, V., and Su, Y. Lifting scheme for lossy attribute encoding in TMC1. Document ISO/IEC JTC1/SC29/WG11 m42640, San Diego, CA, US, 2018.

Morton, G. M. A computer oriented geodetic data base and a new technique in file sequencing. 1966.

Nguyen, D. T. and Kaup, A. Lossless point cloud geometry and attribute compression using a learned conditional probability model. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):4337 4348, 2023.

Nguyen, D. T., Nambiar, K. G., and Kaup, A. Deep probabilistic model for lossless scalable point cloud attribute compression. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1 5. IEEE, 2023.

Pandey, G., Mc Bride, J. R., and Eustice, R. M. Ford campus vision and Li DAR data set. International Journal of Robotics Research, 30(13):1543 1552, 2011.

Raj, T., Hanim Hashim, F., Baseri Huddin, A., Ibrahim, M. F., and Hussain, A. A survey on Li DAR scanning mechanisms. Electronics, 9(5):741, 2020.

S ebastien, L. and Jonathan, T. A point cloud codec for Li DAR data with very low complexity and latency. ISO/IEC JTC1/SC29/WG07 MPEG, 2021.

Tatoglu, A. and Pochiraju, K. Point cloud segmentation with Li DAR reflection intensity behavior. In 2012 IEEE International Conference on Robotics and Automation, pp. 786 790. IEEE, 2012.

Efficient Li DAR Reflectance Compression via Scanning Serialization

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017.

Viswanath, K., Jiang, P., and Saripalli, S. Reflectivity is all you need!: Advancing Li DAR semantic segmentation. ar Xiv preprint ar Xiv:2403.13188, 2024.

Wang, J., Ding, D., and Ma, Z. Lossless point cloud attribute compression using cross-scale, cross-group, and crosscolor prediction. In 2023 Data Compression Conference (DCC), pp. 228 237, 2023.

Wang, J., Xue, R., Li, J., Ding, D., Lin, Y., and Ma, Z. A versatile point cloud compressor using universal multiscale conditional coding part ii: Attribute. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47 (1):252 268, 2025. doi: 10.1109/TPAMI.2024.3462945.

Wang, P.-S. Oct Former: Octree-based transformers for 3D point clouds. ACM Transactions on Graphics (TOG), 42 (4):1 11, 2023.

Wang, Z., Chen, Z., Wu, Y., Zhao, Z., Zhou, L., and Xu, D. Poin Tramba: A hybrid transformer-mamba framework for point cloud analysis. ar Xiv preprint ar Xiv:2405.15463, 2024.

WG 07 MPEG 3D Graphics Coding and Haptics Coding. CTC on AI-based point cloud coding. Output document N01058, ISO/IEC JTC1/SC29/WG7, 148th MPEG meeting, Kemer, November 2024a.

WG 07 MPEG 3D Graphics Coding and Haptics Coding. Common test conditions for G-PCC. Output document N00944, ISO/IEC JTC1/SC29/WG7, 147th MPEG meeting, Sapporo, September 2024b.

Wu, X., Jiang, L., Wang, P.-S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., and Zhao, H. Point Transformer V3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4840 4851, 2024.

You, K., Gao, P., and Ma, Z. Efficient and generic point model for lossless point cloud attribute compression. ar Xiv preprint ar Xiv:2404.06936, 2024.

Zeng, K., Shi, H., Lin, J., Li, S., Cheng, J., Wang, K., Li, Z., and Yang, K. Mamba MOS: Li DAR-based 3D moving object segmentation with motion-aware state space model. In Proceedings of the 32nd ACM International Conference on Multimedia, pp. 1505 1513, 2024.

Zhang, G., Fan, L., He, C., Lei, Z., Zhang, Z., and Zhang, L. Voxel Mamba: Group-free state space models for point cloud based 3D object detection. ar Xiv preprint ar Xiv:2406.10700, 2024a.

Zhang, J., Wang, J., Ding, D., and Ma, Z. Scalable point cloud attribute compression. IEEE Transactions on Multimedia, 2023.

Zhang, T., Yuan, H., Qi, L., Zhanng, J., Zhou, Q., Ji, S., Yan, S., and Li, X. Point Cloud Mamba: Point cloud learning via state space model. AAAI, 2025.

Zhang, W., Yang, F., Xu, Y., and Preda, M. Standardization status of MPEG geometry-based point cloud compression (G-PCC) edition 2. In 2024 Picture Coding Symposium (PCS), pp. 1 5. IEEE, 2024b.

Efficient Li DAR Reflectance Compression via Scanning Serialization

A. Testing Dataset Details

Three widely accepted datasets, including KITTI, Ford, and nu Scenes, are used as testing datasets in this work. We visualize typical samples from these three datasets in Figure 1. Accordingly, Table 1 provides their detailed information.

(c) nu Scenes

Figure 1. Visualization of samples in KITTI, Ford, and nu Scenes Li DAR point cloud datasets. The color indicates the value of reflectance, ranging from blue (low) to red (high).

Table 1. Detailed information of testing datasets

CLASS SEQ. FRAMES REFLECTANCE VALUE POINTS PER FRAME

11 90 0-99 124,130 12 90 0-99 110,921 13 90 0-99 111,997 14 90 0-99 124,628 15 90 0-99 122,315 16 90 0-99 125,622 17 90 0-99 113,687 18 90 0-99 124,873 19 90 0-99 121,239 20 90 0-99 116,382 21 90 0-99 123,307

FORD 02 1500 0-255 83,834 03 1500 0-255 84,063

01 90 0-255 29,606 02 90 0-255 29,568 03 90 0-255 27,560 04 90 0-255 26,854 05 90 0-255 30,478

B. Complexity and Analysis

All the computational complexity reported in the following are derived on our experiment platform equipped with an NVIDIA RTX 4090 GPU, an Intel Core i9-13900K CPU, and 64GB of memory.

B.1. Ser Li C Runtime Analysis

The coding process of Ser Li C consists of three stages: context construction (CC, executed on CPU), neural network (NN, executed on GPU), and arithmetic coding (AC, executed on CPU). These three stages work following the proposed Dual Parallelization mechanism when processing a frame, including the sequence level and the window level, as illustrated in Figure 2(a). Each sequence is divided into windows of equal size and all windows are processed in parallel.

Based on the above Dual Parallelization within a frame, we offer two ways to implement Ser Li C across point cloud frames:

Efficient Li DAR Reflectance Compression via Scanning Serialization

Window 1D point sequences

W0 W0 W0 W0

S0 S1 S2 S3

W1 W1 W1 W1

(a) Details of Dual Parallelization in a Frame

0 0.3 0.6 0.9 1.2 1.5 1.8 2.1 2.4 2.7

Elapsed Time (s)

(b) Example of w/o Frame-level Pipeline

0 0.3 0.6 0.9 1.2 1.5 1.8 2.1 2.4

Elapsed Time (s)

(c) Example of w/ Frame-level Pipeline

Figure 2. (a) Proposed Dual Parallelization within a frame. (b) Sequential implementation across frames. (c) Frame-level pipeline across frames. Si indicates the point sequence. Wi denotes the window. Fi denote the i-th point cloud frame. We use Ser Li C (light) to show the examples in (b) and (c).

Frame-level sequential implementation. We implemented a frame-level sequential version of Ser Li C, running point cloud frames sequentially on our platform, without any pipeline structure, as illustrated in Figure 2(b). As reported in Table 2 and Figure 3(a)-(b), our Ser Li C attains 0.18/0.23 seconds per frame for encoding/decoding, i.e., >5/4 fps (frames per second). Ser Li C (light) achieves a faster speed of >10 fps. Even using the sequential version, our Ser Li C achieves state-of-the-art processing speed. Its light version can meet typical demands in applications such as conventional autonomous driving, Li DAR-based mapping, and geospatial surveying.

Frame-level pipeline implementation. To accelerate processing, we arrange CC, NN, and AC on three pipeline stages for frame-level parallelism, as illustrated in Figure 2(c). As presented in Table 2 and Figure 3(c)-(d), using the three-stage pipelining across frames, Ser Li C runs approximately at a speed of 0.09/0.13 seconds per frame for encoding/decoding, i.e., 11/7.7 fps. The lightweight version of Ser Li C, Ser Li C (light), is even faster, achieving over 30 fps, meeting the real-time requirements of many Li DAR-related applications like real-time obstacle detection in autonomous driving, virtual and augmented reality (VR/AR), and high-fidelity 3D reconstruction.

Table 2. Analysis of encoding/decoding time (seconds per frame), including context construction (CC) latency, neural network (NN) latency, and arithmetic coding (AC) latency

METHOD CC (S) NN (S) AC (S) TOTAL (S) W/ PIPELINE TOTAL (S) W/O PIPELINE

SERLIC 0.03/0.03 0.09/0.13 0.06/0.07 0.09/0.13 0.18/0.23 SERLIC (LIGHT) 0.03/0.03 0.02/0.03 0.03/0.03 0.03/0.03 0.08/0.09

Table 3. Quantitative compression results of Attention-based and Mamba-based implementations

CLASS BPP CR GAIN ATTENTION-BASED MAMBA-BASED VS. ATTENTION-BASED

KITTI 3.80 3.64 -4.21%

Efficient Li DAR Reflectance Compression via Scanning Serialization

0 5 10 15 20 25 30 35 40 45 50

Unicorn (< 1 FPS)

Ser Li C (~ 12 FPS)

Ser Li C (~ 5 FPS)

G-PCC (Predlift) (~ 1 FPS)

(a) Encoding Speed w/o Frame-level Pipeline

0 5 10 15 20 25 30 35 40 45 50

Unicorn (< 1 FPS)

Ser Li C (~ 11 FPS)

Ser Li C (~ 4 FPS)

G-PCC (Predlift) (~ 1 FPS)

(b) Decoding Speed w/o Frame-level Pipeline

0 5 10 15 20 25 30 35 40 45 50

Unicorn (< 1 FPS)

Ser Li C (~ 33 FPS)

Ser Li C (~ 11 FPS)

G-PCC (Predlift) (~ 1 FPS)

(c) Encoding Speed w/ Frame-level Pipeline

0 5 10 15 20 25 30 35 40 45 50

Unicorn (< 1 FPS)

Ser Li C (~ 33 FPS)

Ser Li C (~ 7 FPS)

G-PCC (Predlift) (~ 1 FPS)

(d) Decoding Speed w/ Frame-level Pipeline

Figure 3. The coding speed of Ser Li C w/o and w/ frame-level pipeline implementation.

Table 4. Computational complexity of Attention-based Ser Li C at different window sizes. Maximum GPU memory (Mem.) and encoding/decoding time (seconds per frame) are analyzed. CC, NN, and AC refer to context construction, neural network, and arithmetic coding latencies, respectively.

SIZE MEM. (GB) CC (S) NN (S) AC (S) TOTAL(S)

64 1.31 0.03/0.03 0.03/2.14 0.04/0.04 0.10/2.21 128 1.46 0.03/0.03 0.04/5.02 0.06/0.07 0.13/5.12 256 2.03 0.03/0.03 0.05/13.2 0.09/0.09 0.17/13.3 512 3.23 0.03/0.03 0.09/39.8 0.18/0.18 0.30/40.0 1024 5.84 0.03/0.03 0.15/137 0.32/0.33 0.50/138

B.2. Attention Performance and Complexity

As described in the main manuscript, Attention can also be used in Ser Li C for context modeling. However, compared to Attention, Mamba is more suited to our framework due to its inherent ability to capture point dependencies in a sequence.

To support our claim above, we provide a detailed analysis of both the coding performance and computational complexity when using Attention. This allows readers to better understand the comparative impact of Attention and Mamba on the overall system performance.

Implementation Details. Given the parallel window size W, the computational complexity of self-attention mechanism is Θ(W 2) due to the matrix-based multiplication computation between Q, K, and V in self-attention. As we use the autoregressive coding method, a causal mask with the size of W W (as indicated in Figure 3b of our main manuscript) is applied to ensure that each token in the sequence only uses its preceding tokens (its following tokens are unavailable as they are not coded yet) for context modeling. During the encoding process, since all points are known to the encoder, the model can process all tokens in parallel, and the complexity is accordingly Θ(W 2). However, during decoding, to generate an output sequence consistent with the encoding, the model must decode each point sequentially in a window W. As such, the attention mechanism results in a cubic time complexity Θ(W 3) in decoding.

By contrast, the autoregressive nature in Mamba naturally aligns with our Ser Li C in the 1D point sequence. Its complexity is linear to the window size, i.e., Θ(W). Therefore, using Mamba is more efficient in our implementation.

Coding Performance. Table 3 reports the compression performance of Attention on the KITTI dataset when using the same window size (128) and parameter settings as Mamba. It is observed that Mamba even attains better results than Attention on KITTI, gaining 4.21% on average.

Complexity. Table 4 details the computational complexity of using the Attention block instead of Mamba in Ser Li C. For fair comparisons, we list the runtime of sequential implementation (w/o frame pipelining). In comparison with Table 5 in our

Efficient Li DAR Reflectance Compression via Scanning Serialization

main manuscript where the computational complexity of Mamba is presented, we observe that using Attention requires not only much higher GPU memory (1.46 GB vs. 0.41 GB when window size is 128) but also longer processing time (0.13/5.12 vs. 0.16/0.18 seconds per frame when window size is 128). Obviously, for the encoding time, Attention is comparable to Mamba. But for the decoding time, Attention is much longer, 28 of Mamba.

B.3. L3C2 Analysis and Complexity

Limited by pages, the details of the L3C2 method are not provided in our main manuscript. In the following section, we will elaborate on L3C2 and its complexity for comparison with our Ser Li C.

L3C2 Introduction. MPEG introduced L3C2 (S ebastien & Jonathan, 2021) in 2021, specifically tailored for the compression of Li DAR point clouds. However, its dependence on detailed sensor configurations (e.g., the precise pitch angles of individual lasers) limits its applicability on various datasets. Specifically, L3C2 requires the following inputs (refer to Figure 4):

(1) num Lasers: the number of laser scan lines from the Li DAR;

(2) lasers Theta: the tangent of the pitch angle for each line, where negative values indicate a scan direction towards the ground, and positive values indicate a scan towards the sky;

(3) lasers Z: the distance from the laser head to the Li DAR center for each line, measured in millimeters;

(4) lasers Num Phi Per Turn: the number of points each line can scan in one full rotation.

For readers convenience, we capture the configuration of L3C2 codec in Figure 4 to demonstrate the parameters required by L3C2. Since MPEG provides detailed Li DAR configuration parameters only for the Ford dataset, we are unable to test on KITTI and nu Scenes in this work. Moreover, these parameters are only applicable to raw, unprocessed Li DAR point cloud data. However, the point cloud samples in nu Scenes and KITTI datasets have undergone preprocessing before release. As a result, even if their Li DAR parameters could be derived, they would not be compatible with these preprocessed samples.

Figure 4. Configuration of L3C2 codec, where four parameters are obtained from raw Li DAR physical parameters.

Table 5. Runtime comparison against L3C2 on the Ford dataset

METHOD ENC. DEC.

L3C2 0.11S 0.06S SERLIC 0.19S 0.22S SERLIC (LIGHT) 0.08S 0.08S

Complexity. In Table 2 of our main manuscript, we compare the coding performance of Ser Li C with that of L3C2. Furthermore, we compare their runtime in Table 5. It is observed that our Ser Li C (light) runs slightly faster than L3C2 in encoding process, while slightly lower in decoding. Notice that L3C2 is implemented using C language and well optimized on CPU, whereas our Ser Li C uses Python language and executes on both CPU and GPU. Thus, this comparison just serves as a reference for intuitive observation about the runtime of both methods.

C. G-PCC Configuration

G-PCC offers an optional configuration, called the Angular mode, which includes detailed Li DAR parameters similar to those used in L3C2. Since these parameters are only available on the Ford dataset, we enable the Angular mode only when testing on the Ford dataset, and disable it for KITTI and nu Scenes datasets.

Efficient Li DAR Reflectance Compression via Scanning Serialization

Table 6. Detection results using the classical detection models w/ and w/o reflectance on the KITTI dataset

METHODS CAR PED. CYC. MAP EASY MOD. HARD EASY MOD. HARD EASY MOD. HARD MOD.

POINTPILLAR (W/ R) 87.75 78.40 75.18 57.30 51.41 46.87 81.57 62.81 58.83 64.21 POINTPILLAR (W/O R) 83.85 74.21 70.36 21.56 14.08 12.83 47.79 34.28 32.13 40.86

SECOND (W/ R) 90.55 81.61 78.61 55.95 51.15 46.17 82.97 66.74 62.78 66.50 SECOND (W/O R) 87.87 78.92 75.58 40.46 35.66 31.88 70.42 50.64 47.70 55.07

POINTRCNN (W/ R) 91.47 80.54 78.05 62.96 55.04 48.56 89.17 70.89 65.64 68.82 POINTRCNN (W/O R) 88.03 77.00 72.85 42.90 35.28 30.68 56.58 42.51 40.07 51.60

D. Downstream Tasks

In our introduction, we mention that ...our experiments show that removing the reflectance causes a dramatic drop in pedestrian (and cyclist) detection accuracy incorporating the pre-trained Point Pillar (Lang et al., 2019) detection model, having AP (average precision) from 51.4 (62.8) to 14.1 (34.3) on the KITTI dataset (see Figure 1)... To support our statement, we offer our full results on the KITTI dataset in Table 6.

The results in Table 6 provide a detailed comparison of the detection performance of three classical models Point Pillar, SECOND, and Point RCNN with and without reflectance data. Note that our experiments are conducted on pre-trained models. The table shows the Average Precision (AP) for three categories: Car, Pedestrian (Ped.), and Cyclist (Cyc.), across different difficulty levels (Easy, Moderate, Hard), as well as the mean Average Precision (m AP) for the Moderate category.

It is observed that, reflectance data is essential for maintaining high accuracy and robustness in object detection tasks, particularly for pedestrians and cyclists. The substantial drop in AP for these categories underscores the importance of preserving reflectance information in point cloud data for downstream applications.

(a) Overall

(b) Zoom-in (region)

(c) Zoom-in (line)

Figure 5. Intuitive observation on reflectance characteristics in KITTI. (a) shows a holistic perspective. (b) shows zoom-in the region details. (c) further zooms in the reflectance of each sensor line for observation. The color indicates the value of reflectance, ranging from blue (low) to red (high).

Table 7. Ablation study on the sequence-level parallelization: exploiting correlations across sequences. Intra indicates intra-sequence coding, while Inter (1) and Inter (3) represent inter-sequence coding based on the preceding one or three sequences, respectively.

BPP MEM. (GB) NN (S) TOTAL (S)

INTRA 3.64 0.41 0.09/0.13 0.18/0.23 INTER (1) 3.64 0.04 5.04/5.82 5.20/9.84 INTER (3) 3.63 0.04 5.25/6.18 5.48/10.16

E. More Ablation Studies

Sequence Parallelization. In Section 3.4 Dual Parallelization, we propose the dual parallelization scheme. including sequence level and window level, to accelerate coding speed of Ser Li C within a frame. The ablation of window-level parallelization is presented in Section 4.5 Window-based Parallelization. Limited by page number, we depict the ablation study of Sequence-level Parallelization here.

Efficient Li DAR Reflectance Compression via Scanning Serialization

On the other hand, in Section 3.4 Dual Parallelization, we state that Ser Li C processes point sequences in parallel, disregarding correlations across sequences. This is justified, as point sequences are scanned by separate laser sensor rotations, inherently lacking strong inter-sequence dependencies. As a result, sequence parallelization does not compromise coding efficiency but greatly enhances processing speed. Our ablation study here supports our statement above.

Figure 5 visualizes the reflectance of a point cloud frame. For better observation, we color the reflectance values using different colors. It is clear that reflectance values are close to each other in a scan line while different to some extent across scan lines. This intuitively reveals that using point correlations within a scan line leads to better performance. Furthermore, we conduct experiments which exploit correlations not only across points in a sequence but also across sequences for reflectance compression. In this way, the sequence-level parallelization is disabled. That is, we apply the autoregressive coding both across sequences (called inter-sequence coding in the following) and within a sequence (called intra-sequence coding in the following).

Compression Performance. Table 7 presents the experimental results. When leveraging correlations between two sequences, i.e., using the previous sequence for context modeling while encoding the current one, no gain is attained. The same holds when utilizing the previous three sequences. These findings is consistent with our observation that points are more strongly correlated within a single sequence. This further validates the rationale behind our sequence-level parallelization approach, which effectively maintains high coding performance while keeping the complexity low.

Computational Complexity. As the inter-sequence coding disable sequence-level parallelism, its memory cost is reduced. As reported in Table 7, the memory usage of inter-sequence coding is only 0.04 GB. However, the runtime is accordingly remarkably increased due to the autoregressive coding across sequences. For example, the total encoding/decoding time is 5.20/9.84 seconds per frame when using the previous sequence as context of the current sequence. When using three sequences, the coding time is even longer (5.48/10.16) as longer time is required by the context construction stage.