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

Pedestrian detection #45

Open ben0110 opened 5 years ago

ben0110 commented 5 years ago

Hi,

I'm working on collecting and annotating data with a stereo camera for pedestrian detection and for accelerating the annotation process, I was thinking about using an already trained algorithm for generating annotation boxes proposals and as Stereo-RCNN is on top of the list for Kitti vehicle detection benchmark, wanted to ask if you already tested Stereo-RCNN for pedestrian detection and would it be much overhead/worth it to train it for pedestrian detection.

Best regards, Amine

PeiliangLi commented 5 years ago

This work may not work very well for pedestrian, because we use a "box-shape" to model the object shape. As we mentioned in the paper, this issue can be addressed by leaning general object shape.