Diffusion Models-Based Purification for Common Corruptions on Robust 3D Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• LiDAR data corruption hinders 3D vision tasks like object detection.
• Traditional defenses (e.g., adversarial training) face gradient confusion and limited corruption-specific robustness.

Purpose of the Study:

• To introduce LiDARPure, a novel method for purifying corrupted LiDAR scene data.
• To overcome limitations of adversarial training in handling sparse point clouds and gradient confusion.

Main Methods:

• LiDARPure employs diffusion models for data purification.
• The method divides scenes into voxels for diffusion and reverse diffusion processes.
• Latent geometric features condition the diffusion model's generation process.

Main Results:

• LiDARPure effectively purifies 19 common types of LiDAR data corruption.
• It improves the average precision of 3D object detectors by up to 20% when facing data corruption.
• Performance significantly surpasses existing defense strategies.

Conclusions:

• LiDARPure offers a robust solution for LiDAR data corruption in 3D vision.
• The diffusion model approach provides superior performance compared to traditional methods.