search

Owen-Liuyuxuan / visualDet3D

Official Repo for Ground-aware Monocular 3D Object Detection for Autonomous Driving / YOLOStereo3D: A Step Back to 2D for Efficient Stereo 3D Detection
https://owen-liuyuxuan.github.io/papers_reading_sharing.github.io/3dDetection/GroundAwareConvultion/
Apache License 2.0
361 stars 76 forks source link

Is YOLOStereo3D aware of the stereo baseline? #60

Open ootts opened 2 years ago

ootts commented 2 years ago

Hi, I wonder whether the stereo baseline is used in YOLOStereo3D. If I want to use YOLOStereo3D trained on KITTI to perform object detection in another scenario where the stereo baseline is different from KITTI, will it work? Which part should I modify?

Thanks a lot!

Owen-Liuyuxuan commented 2 years ago

Great question. We do not have the stereo baseline explicitly used in inference. So the network is kind of overfitting to the camera setup of the dataset.

Some other users have reported that YOLOStereo3D does not generalize well when directly moving on to a new personalized camera setup even at 2D detection (but we are not sure about the stereo calibration accuracy of their setups).

From my personal LIMITED experience, I will first check if the 2D detection results transfer well (if it fails here then it is hard to transfer); if so, then I will try to compute a mapping between the predicted depth from the trained network to gt depth at the new scene (like calibrating the scale).