Super Sparse 3D Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• LiDAR-based 3D object detection is crucial for autonomous driving.
• Current detectors struggle with long-range perception due to high computational costs from dense feature maps.

Purpose of the Study:

• To develop efficient long-range 3D object detection methods for autonomous driving.
• To overcome the limitations of dense feature maps in scaling with perception range.

Main Methods:

• Proposed Fully Sparse Detector (FSD) using a sparse voxel encoder and Sparse Instance Recognition (SIR) module.
• Developed FSD++ by leveraging temporal information and residual points to create super sparse input data, reducing redundancy.
• Utilized instance-wise grouping to address center feature missing issues in fully sparse architectures.

Main Results:

• Achieved state-of-the-art performance on the Waymo Open Dataset.
• Demonstrated superior performance in long-range detection scenarios on the Argoverse 2 Dataset (up to 200m range).
• Significantly reduced computational overhead and data redundancy with FSD++.

Conclusions:

• FSD and FSD++ offer efficient and scalable solutions for long-range LiDAR-based 3D object detection.
• The proposed methods significantly advance the capabilities of autonomous driving perception systems.
• Sparse detection architectures are viable and effective for long-range perception tasks.