Not All Points Are Equal: Learning Highly Efficient Point-based Detectors for 3D LiDAR Point Clouds (2203.11139v1)

Abstract: We study the problem of efficient object detection of 3D LiDAR point clouds. To reduce the memory and computational cost, existing point-based pipelines usually adopt task-agnostic random sampling or farthest point sampling to progressively downsample input point clouds, despite the fact that not all points are equally important to the task of object detection. In particular, the foreground points are inherently more important than background points for object detectors. Motivated by this, we propose a highly-efficient single-stage point-based 3D detector in this paper, termed IA-SSD. The key of our approach is to exploit two learnable, task-oriented, instance-aware downsampling strategies to hierarchically select the foreground points belonging to objects of interest. Additionally, we also introduce a contextual centroid perception module to further estimate precise instance centers. Finally, we build our IA-SSD following the encoder-only architecture for efficiency. Extensive experiments conducted on several large-scale detection benchmarks demonstrate the competitive performance of our IA-SSD. Thanks to the low memory footprint and a high degree of parallelism, it achieves a superior speed of 80+ frames-per-second on the KITTI dataset with a single RTX2080Ti GPU. The code is available at \url{https://github.com/yifanzhang713/IA-SSD}.

• The paper introduces novel instance-aware downsampling strategies, combining class-aware and centroid-aware sampling to reduce computational load while preserving key foreground points.
• The contextual centroid perception module enhances localization accuracy by aggregating spatial features around predicted bounding box centers.
• Evaluations on KITTI, Waymo, and ONCE demonstrate real-time detection at over 80 FPS, with robust instance recall even for small objects.

Overview of IA-SSD: Efficient 3D Object Detection for LiDAR Point Clouds

In the context of 3D object detection using LiDAR point clouds, efficient processing is essential due to the voluminous nature of point cloud data and the importance of identifying salient features for reliable detection. The authors present IA-SSD, a novel approach designed to enhance point-based 3D object detection by leveraging task-aware point sampling and contextual centroid perception.

Methodology

The cornerstone of IA-SSD is its introduction of two distinct task-oriented, instance-aware downsampling strategies designed to preserve informative foreground points during the downsampling process. Unlike traditional methods reliant on random or farthest point sampling, these strategies prioritize points likely associated with objects of interest, thereby reducing computational load while maintaining detection accuracy.

1. Class-aware and Centroid-aware Sampling:
   • Class-aware Sampling: Utilizes semantic priors by incorporating a parallel MLP layer to predict the semantic category probability of each point, thus informing downsampling decisions.
   • Centroid-aware Sampling: Focuses on proximity to instance centroids, with point weights influenced by the spatial closeness to predicted object centers.
2. Contextual Centroid Perception Module:
   • Enhances accuracy by exploiting contextual information around predicted bounding box centers, achieving better localization through aggregation of spatial features around detected instances.

Results and Evaluation

IA-SSD demonstrates competitive performance across several large-scale benchmarks, including KITTI, Waymo, and ONCE datasets. Notably, it achieves a frame rate of over 80 FPS on an RTX2080Ti GPU, highlighting its suitability for real-time applications. The strong instance recall rates presented in the ablation studies underscore the effectiveness of the proposed instance-aware sampling in maintaining performance even for small objects like pedestrians and cyclists.

Implications

From a practical standpoint, IA-SSD presents a scalable and efficient solution for autonomous systems requiring fast and accurate 3D object detection. The ability to detect multiple object categories within a single model enhances deployment flexibility, simplifying system integration.

Future Directions

Potential future developments could include refining the integration of multi-scale feature aggregation to handle varying object sizes more robustly, particularly enhancing performance for large vehicles in complex environments such as those found in autonomous driving. Additionally, further exploring the balance between computational efficiency and detection performance could lead to broader applications beyond autonomous vehicles, such as robotics and augmented reality.

In summary, IA-SSD signifies a meaningful step toward efficient and accurate 3D object detection in LiDAR point clouds, offering insights into both methodological advancements and real-world applicability.