Efficient LiDAR Reflectance Compression via Scanning Serialization

TL;DR

SerLiC introduces a scan-order serialization framework for LiDAR reflectance compression, converting 3D point clouds into 1D sequences that align with LiDAR scanning physics. It uses physics-informed contextual tokens and a Selective State Space Model (Mamba) for autoregressive coding, supported by dual sequence/window parallelization to achieve linear-time complexity and real-time performance. The approach delivers over 2× volume reduction and up to 22% better bit-rate than prior state-of-the-art methods (e.g., Unicorn) with only about 2% of their parameters, and a light variant reaches >30 fps on typical hardware. These results demonstrate strong compression efficiency, ultra-low resource demands, and robustness across KITTI, Ford, nuScenes, and even non-rotational datasets, highlighting practical applicability for real-world LiDAR systems.

Abstract

Reflectance attributes in LiDAR point clouds provide essential information for downstream tasks but remain underexplored in neural compression methods. To address this, we introduce SerLiC, a serialization-based neural compression framework to fully exploit the intrinsic characteristics of LiDAR reflectance. SerLiC first transforms 3D LiDAR 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 SerLiC attains over 2x 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 SerLiC achieves > 10 fps (frames per second) with just 111K parameters, which is attractive for real-world applications.

Figures (9)

• Figure 1: (a-c) Reflectance plays an indispensable role in downstream tasks such as 3D object detection; (d) The proposed SerLiC establishes a new state-of-the-art by outperforming the latest compression standard G-PCC zhang2024standardization and the learning-based work Unicorn on widely accepted datasets KITTI Behley2019KITTI, Ford ford, and nuScenes caesar2020nuscenes (the number denotes sequence number); (e) SerLiC achieves ultra-low coding latency with an exceptionally lightweight model.
• Figure 1: Visualization of samples in KITTI, Ford, and nuScenes LiDAR point cloud datasets. The color indicates the value of reflectance, ranging from blue (low) to red (high).
• Figure 2: SerLiC Framework. The input 3D LiDAR 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 ($\mathbf{F}_{i}^{pos}$), radial distance ($\mathbf{F}_{i}^{\rho}$), and prior reflectance ($\mathbf{F}_{i-1}^{\mathbf{x}}$) as context to generate the probability mass function (PMF) for the reflectance intensity of the target ($i$-th) point.
• Figure 2: (a) Proposed Dual Parallelization within a frame. (b) Sequential implementation across frames. (c) Frame-level pipeline across frames. $\mathrm{S}_i$ indicates the point sequence. $\mathrm{W}_i$ denotes the window. $\mathrm{F}_i$ denote the $i$-th point cloud frame. We use SerLiC (light) to show the examples in (b) and (c).
• Figure 3: Basic Mamba block and Attention block. "LN" refers to Layer Norm; $\sigma$ denotes SiLU activation; $\otimes$ means Hadamard product; $\oplus$ represents element-wise addition.