SCNet: Subdivision Coding Network for Object Detection Based on 3D Point Cloud

Zhiyu Wang, Hao Fu, Li Wang, Liang Xiao, Bin DaiIEEE Access 2019 · DOI

Abstract

This paper presents SCNet (Subdivision Coding Network), a real-time 3D object detector designed for autonomous driving applications. SCNet addresses the computational inefficiency of 3D voxel methods and the information loss inherent in 2D Bird’s Eye View (BEV) projections. By introducing a novel subdivision coding scheme, the network effectively preserves fine-grained spatial details within a 2D convolutional framework. Experimental results on the KITTI benchmark demonstrate that SCNet achieves state-of-the-art accuracy, particularly for vehicle detection, while maintaining a real-time processing speed of approximately 25 FPS.

Introduction

LiDAR sensors are pivotal for 3D perception in autonomous vehicles, providing high-precision depth information. However, the sparse and unordered nature of point clouds makes efficient processing difficult. Current research typically follows two paths: 3D voxelization, which is accurate but computationally prohibitive, and 2D projection, which is fast but suffers from quantization errors that obscure vertical structural details. SCNet bridges this gap by retaining the speed of 2D convolutions while regaining structural nuances through intelligent data rearrangement.

Methodology

The SCNet architecture is composed of three main stages: Point Rearrangement, Stixel Feature Extraction, and a Region Proposal Network (RPN).

The SCNet architecture pipeline, illustrating the flow from raw point cloud to final prediction.

The core innovation lies in the Subdivision Coding process. During Point Rearrangement, individual grid cells are further divided into sub-grids. The distribution of points across these sub-grids is encoded into a uniform feature vector, effectively packing 3D structural information into a format compatible with 2D networks. This process ensures that the "profile" of objects is preserved, which is critical for distinguishing between classes like pedestrians and light poles.

Visualization of subdivision coding patterns (a-d) representing different spatial point distributions within a grid cell.

For feature extraction, the model utilizes depth-wise separable convolutions, significantly reducing parameter count and computational cost. The final stage employs a single-stage RPN that fuses feature maps from three resolutions to ensure robust detection of objects across various scales.

Key Insights and Results

SCNet was evaluated on the KITTI Object Benchmark, focusing on Cars, Pedestrians, and Cyclists. The model achieved a 90.35% Average Precision (AP) for cars in the "Easy" category, outperforming established baselines like VoxelNet and PointPillars. Ablation studies revealed that using absolute coordinates ($x, y, z$) as input features provided superior spatial context compared to relative coordinates. Furthermore, by processing only non-empty grids (which typically account for less than 10% of the space), the network achieves high efficiency, running at ~40ms per frame on a GTX 1080Ti.

Conclusion

SCNet demonstrates a successful approach to real-time 3D object detection by prioritizing data organization over raw model depth. While highly effective for vehicle detection, future work will focus on improving sensitivity to smaller objects like pedestrians and integrating multi-sensor fusion to overcome texture-based classification challenges.

Summary updated

BibTeX
@article{wang2019scnet,
  author = {Zhiyu Wang and Hao Fu and Li Wang and Liang Xiao and Bin Dai},
  title = {{SCNet: Subdivision Coding Network for Object Detection Based on 3D Point Cloud}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2937676},
}
Figure