eGAC3D: enhancing depth adaptive convolution and depth estimation for monocular 3D object pose detection.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• High-precision 3D LiDAR sensors are costly for 3D object detection.
• Monocular 3D object detection methods aim to reduce costs by using single cameras.
• Previous work introduced GAC3D, utilizing geometric constraints and adaptive depth convolution for improved monocular 3D detection.

Purpose of the Study:

• To present eGAC3D, an improved architecture for monocular 3D object detection.
• To enhance detection accuracy and reduce inference time compared to existing methods.
• To demonstrate the feasibility of deploying advanced monocular 3D detection on embedded platforms.

Main Methods:

• Developed eGAC3D with revised depth adaptive convolution and variant guidance.
• Incorporated pixel adaptive convolution to leverage depth maps for detection heads, eliminating external depth estimators.
• Optimized the eGAC3D framework for embedded platforms with low-cost GPUs.

Main Results:

• eGAC3D demonstrated superior accuracy and faster inference times than GAC3D and other monocular methods on the KITTI benchmark.
• The framework achieved nearly real-time performance on an embedded platform (Jetson Xavier NX) with modest hardware.
• This represents the first monocular 3D detection framework developed and optimized for embedded devices.

Conclusions:

• eGAC3D offers a significant advancement in monocular 3D object detection, balancing accuracy and efficiency.
• The method provides a cost-effective solution for real-time 3D perception on resource-constrained embedded systems.
• This work paves the way for deploying sophisticated 3D object detection in applications limited by cost and computational power.