OBMO: One Bounding Box Multiple Objects for Monocular 3D Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving

Background:

• Monocular 3D object detection offers a simpler alternative to multi-sensor systems but lags behind LiDAR-based methods.
• Depth ambiguity in monocular images, where objects at different depths share similar 2D features, hinders accurate depth estimation and training stability.

Purpose of the Study:

• To address the depth ambiguity challenge in monocular 3D object detection.
• To improve the accuracy and stability of monocular 3D object detection systems.

Main Methods:

• Introduced a plug-and-play module named One Bounding Box Multiple Objects (OBMO).
• Generated pseudo-3D bounding box labels by shifting original boxes along the viewing frustum.
• Developed two label scoring strategies to ensure the quality and reasonableness of pseudo-labels.

Main Results:

• The OBMO module significantly enhances training stability by utilizing soft pseudo-labels with quality scores.
• Achieved substantial performance improvements on KITTI and Waymo benchmarks, outperforming state-of-the-art monocular 3D detectors.
• Demonstrated notable gains in both Bird's-Eye View (BEV) and 3D mean Average Precision (mAP).

Conclusions:

• The proposed OBMO module effectively mitigates depth ambiguity in monocular 3D object detection.
• This approach offers a simple yet powerful method to boost the performance of existing monocular detectors.
• The findings suggest a promising direction for advancing monocular 3D object detection capabilities.