MonoEF: Extrinsic Parameter Free Monocular 3D Object Detection.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving Systems

Background:

• Monocular 3D object detection is crucial for autonomous driving but struggles with variations in ego-car pose relative to the ground plane.
• Existing methods often ignore camera pose information, making them vulnerable to extrinsic parameter changes and failing in real-world scenarios like uneven roads.

Purpose of the Study:

• To develop a novel method for monocular 3D object detection that is robust to camera extrinsic parameter variations.
• To improve the performance of 3D object detectors in realistic autonomous driving conditions with fluctuating road surfaces.

Main Methods:

• The proposed framework predicts camera extrinsic parameters by detecting the vanishing point and horizon change.
• A feature rectification converter is employed to address perturbative features in the latent space, making the detector independent of extrinsic parameter variations.

Main Results:

• The novel 3D detector demonstrates independence from extrinsic parameter variations, achieving accurate results in challenging real-world scenarios.
• Experiments on the KITTI 3D and nuScenes datasets show superior performance compared to state-of-the-art methods by a significant margin.

Conclusions:

• The proposed method effectively handles camera pose changes, a common issue in autonomous driving.
• This approach offers a robust solution for monocular 3D object detection, particularly in environments with uneven terrain where other methods falter.