Vertex points are not enough: Monocular 3D object detection via intra- and inter-plane constraints

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Current 3D object detection methods primarily focus on rigid objects like cars.
• Detection of deformable objects, such as cyclists, using monocular images remains a significant challenge.

Purpose of the Study:

• To develop a novel 3D monocular object detection method.
• To enhance detection accuracy for objects with significant deformation, specifically cyclists.
• To improve depth location prediction in monocular 3D object detection.

Main Methods:

• Introduction of geometric constraints for the object's 3D bounding box plane.
• Incorporation of intra-plane constraints for keypoint regression, ensuring errors remain within the projection plane.
• Utilizing prior knowledge of inter-plane geometry to optimize keypoint regression and depth prediction.

Main Results:

• The proposed method demonstrates superior performance on the cyclist detection class compared to state-of-the-art techniques.
• Achieves competitive results in real-time monocular 3D object detection.
• Significantly improves the accuracy of depth location prediction for deformable objects.

Conclusions:

• The novel geometric constraint approach effectively addresses limitations in detecting deformable objects from monocular images.
• The method offers a promising solution for accurate and real-time 3D cyclist detection.
• This work advances the field of monocular 3D object detection by incorporating object-specific geometric priors.