Exploiting Ground Depth Estimation for Mobile Monocular 3D Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Monocular 3D object detection is vital for mobile applications like autonomous vehicles and drones.
• Challenges include near-far depth disparity and dynamic camera poses, impacting accuracy, especially for distant objects.

Purpose of the Study:

• To propose MoGDE, a new framework enhancing monocular 3D object detection by utilizing ground depth information.
• To improve the accuracy and robustness of 3D object detection in mobile applications.

Main Methods:

• MoGDE estimates ground depth and uses it to guide the Mono3D framework.
• A pose detection network estimates camera orientation, creating a pixel-level ground depth feature map.
• An RGB-D feature fusion network with transformer architecture integrates ground depth information.

Main Results:

• MoGDE significantly improves accuracy and robustness for both near and far objects.
• The framework achieves state-of-the-art performance on the KITTI dataset.
• MoGDE ranks first among pure image-based methods on the KITTI 3D benchmark.

Conclusions:

• Leveraging ground depth is an effective strategy for improving monocular 3D object detection.
• MoGDE offers a robust and accurate solution for 3D perception in mobile robotic systems.