Doubt about training with dimension and orientation

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

Doubt about training with dimension and orientation #46

Open liangkangzheng opened 5 years ago

liangkangzheng commented 5 years ago

Hi Peiliang,

recently I've studied your code about Stereo RCNN. It's a great pleasure to study such a good method in image-based object detection. During studying the code, I'm not quite sure about some of them and hope you could explain them...

During training process, the outside (and inside) bounding box weights for dimension orientation was not implemented as other did. why's that?

PeiliangLi commented 5 years ago

The inside and outside ws is for masking the negative roi and weighting the different regression term. Thanks for pointing out this. It might be a bug. I will check that.