- pytorch>=1.2
- opencv-python
- torchvision>=0.4
Boxes are in different scale. Anchor points with more than one matched gt box are ignored.
green box --> gt box
blue/red point --> positive sample
The positive sample sampling of the fifth branch is shown below.
- regression target
- add BN
- more classes
I have trained the model on a safety helmet dataset without any augmentation. It seems to be working well.
tensorboard
test_result