DOtA++: Unsupervisely and Collaboratively Detect Objects From Multi-Agent Observations With Multi-Modal Prior...

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• Multi-agent collaboration enhances perception in autonomous driving.
• Manual annotation for training collaborative detectors becomes increasingly burdensome with more agents.

Purpose of the Study:

• To develop an unsupervised method for object detection from multi-agent LiDAR scans without external labels.
• To improve upon the initial method with enhanced constraints and multi-modal data.

Main Methods:

• DOtA: Generates preliminary labels via an initial detector trained on shared agent information, then refines labels using physical rule constraints.
• DOtA++: Builds on DOtA, incorporating composite prior constraints, image data for multi-agent observation consistency, and point cloud geometric distribution constraints.

Main Results:

• DOtA and DOtA++ effectively detect objects in scenes without manual annotations.
• DOtA++ achieved a 10.7% mAP improvement over traditional unsupervised methods on the V2X-R dataset.

Conclusions:

• Unsupervised learning is a viable approach for training collaborative object detectors in autonomous driving.
• DOtA++ demonstrates significant performance gains through the integration of diverse constraints and multi-modal data.