MonoAux: Fully Exploiting Auxiliary Information and Uncertainty for Monocular 3D Object Detection.

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Machine Learning

Background:

• Monocular 3D object detection is crucial for autonomous driving but challenging due to lack of depth information.
• Existing methods often fail to fully utilize limited monocular data, lacking uncertainty analysis and post-processing.
• This leads to suboptimal performance in precise 3D object localization from single images.

Purpose of the Study:

• To develop a comprehensive framework for monocular 3D object detection that maximizes information extraction.
• To incorporate diverse depth estimation techniques and uncertainty analysis to improve localization accuracy.
• To address challenges in multi-task training and information scarcity in monocular detection.

Main Methods:

• Mining intrinsic image information for augmented supervision to overcome data limitations.
• Recovering multiple depth values from visual heights for robust depth estimation.
• Employing an uncertainty fusion process to determine final depth and confidence, reducing inference errors.
• Implementing an adaptive training strategy with measurement indicators for dynamic task weight adjustment.

Main Results:

• The proposed framework demonstrates enhanced performance on KITTI and Waymo datasets across various difficulty levels.
• The method consistently outperforms the original framework in monocular 3D detection accuracy.
• Real-time efficiency is maintained while achieving superior results.

Conclusions:

• The developed framework effectively addresses information scarcity and uncertainty in monocular 3D object detection.
• The novel depth estimation and uncertainty fusion techniques significantly reduce inference errors.
• The adaptive training strategy optimizes multi-task learning for improved overall performance in autonomous driving applications.