Weakly Supervised Monocular 3D Object Detection by Spatial-Temporal View Consistency.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Monocular 3D object detection is vital for autonomous vehicles but suffers from a training-inference data discrepancy.
• Current methods require 3D ground truths from LiDAR during training, hindering the use of real-world driving data.
• This gap prevents continuous improvement of 3D object detectors in production vehicles.

Purpose of the Study:

• To develop a weakly-supervised approach for monocular 3D object detection that eliminates the need for 3D ground truths.
• To establish a connected data loop for continuous improvement of 3D object detection models.
• To enhance the robustness and accuracy of 3D object detectors by utilizing only 2D labels.

Main Methods:

• A weakly-supervised learning framework utilizing only 2D labels for training monocular 3D object detectors.
• Implementation of spatial view consistency through projection and multi-view techniques for optimizing object location and size.
• Leveraging temporal view consistency and introducing temporal movement consistency to address dynamic scenes and improve 3D bounding box prediction.

Main Results:

• Achieved performance comparable to fully supervised methods using only 2D ground truths.
• Demonstrated the effectiveness of spatial and temporal view consistency in regulating 3D bounding box predictions.
• Showcased the method's utility as a pre-training strategy, yielding significant gains when fine-tuned with limited 3D data.

Conclusions:

• Weakly-supervised monocular 3D object detection using 2D labels is feasible and effective.
• The proposed view consistency techniques are crucial for accurate 3D perception without 3D ground truth.
• This approach offers a viable solution for data-efficient training and continuous learning in autonomous driving systems.