search

HKUST-Aerial-Robotics / Stereo-RCNN

Code for 'Stereo R-CNN based 3D Object Detection for Autonomous Driving' (CVPR 2019)
MIT License
692 stars 177 forks source link

low accuracy #53

Closed yxy1995123 closed 5 years ago

yxy1995123 commented 5 years ago

I trained the mono branch and tested without dense alignment.The result is bad. bbox AP:92.51, 82.42, 67.09 bev AP:10.17, 6.37, 6.75 3d AP:6.79, 4.92, 3.72 aos AP:90.87, 80.31, 64.89 Is it a normal situation? if not,can you give me some advice?

PeiliangLi commented 5 years ago

Yes, it's normal. In the mono branch, if you don't use the dense alignment, the 3D box is calculated from pure 2d bounding box, which can not guarantee the 3d accuracy.

ianw4tm commented 4 years ago

Did you only use one image for everything in the Mono branch?