About the Vt constraint in box estimator

HKUST-Aerial-Robotics / Stereo-RCNN

Code for 'Stereo R-CNN based 3D Object Detection for Autonomous Driving' (CVPR 2019)

MIT License

690 stars 177 forks source link

About the Vt constraint in box estimator #38

Closed chenchr closed 5 years ago

chenchr commented 5 years ago

Hello. Thanks for the code. I have a question about the Vt (top edge of left 2D box) constraint. It seems that this implementation hardcodes the vertex that should project to the top edge of left 2D box. Maybe it's based on the assumption that the camera is able to see the top plane of 3D box ? However, when the top plane of 3D box is invisible, this specific vertex seems depends on the observation angle and viewpoint as well.

PeiliangLi commented 5 years ago

Basically, the top plane of the 3D box will be visible in driving scenarios. In other scenarios (indoor scene, etc), the top plane of the 3D box might be invisible if the camera has 3-DoF rotation, where we need 3-axis viewpoints estimation to determine the 3D-2D relations.