Aayushktyagi
Aayushktyagi
Check the tracker object and see if trackers are being poped from it.
Turning the shuffle parameter off in the dataloader helped. [Link](https://stackoverflow.com/a/68621335/8924426)
Yes faced the same issue while training on a custom dataset with a single class.
You can copy the PointRend from detectron2 repo and place it in the BoxTeacher/projects/
I think this might help: https://www.kaggle.com/code/hikmatullahmohammadi/road-lane-line-tusimple-dataset-preparation .
pip install numpy==1.19 helps here.
I also observed the same trend. For Boxlevelset and box2mask, performance on single class is very poor. For Boxinst its decent. @VikasRajashekar did you try BoxLevelset?
@imashwini09 I got the same issue. try commenting out {"class_name": "albumentations.ShiftScaleRotate", "params": {"p": 0.5} } Sometimes due to ShiftScaleRotate transform returns empty list of boxes []. Thanks