Towards a Weakly Supervised Framework for 3D Point Cloud Object Detection and Annotation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Manual annotation of LiDAR point clouds for 3D object detection is labor-intensive and expensive.
• Existing fully supervised methods require extensive, precise annotations, hindering practical application.

Purpose of the Study:

• To develop a weakly supervised framework for training high-quality 3D object detectors with reduced annotation effort.
• To enable learning 3D detection from a limited number of weakly annotated examples.

Main Methods:

• A two-stage architecture: Stage-1 generates object proposals using Bird's-Eye View (BEV) center-click annotations, and Stage-2 refines predictions in a coarse-to-fine manner.
• Utilizes a novel BEV center-click annotation strategy and incomplete supervision for training.

Main Results:

• Achieves 86-97% of fully supervised performance on the KITTI dataset using significantly fewer annotations (500 scenes, 534 instances vs. 3,712 scenes, 15,654 instances).
• The trained model functions as an effective 3D object annotator, generating data that trains detectors to over 95% of original performance.
• Demonstrates potential for performance improvement with increased training data.

Conclusions:

• The proposed weakly supervised framework drastically reduces annotation costs for 3D object detection.
• The method is practical, offering automatic and human-in-the-loop annotation modes, and can serve as a pre-training tool.