The variant is named align_variant.yaml.
Main features:
1)Half Instance Normalization adopted.
2)Branch loss weight rebalanced to 1, 4, 16 for faster converge and stabler training.
3)UDIS's backbone used instead of YOLO's.
Other modifications:
1)Bugs fixed in loading image masks, the postfix ".png" is changed to ".jpg".
2)An additional option for training named "loss-weight" to change the aforementioned branch loss weight.
3)Some modifications in yolo.py and common.py to adapt to the new variant.
The variant is named align_variant.yaml. Main features: 1)Half Instance Normalization adopted. 2)Branch loss weight rebalanced to 1, 4, 16 for faster converge and stabler training. 3)UDIS's backbone used instead of YOLO's.
Other modifications: 1)Bugs fixed in loading image masks, the postfix ".png" is changed to ".jpg". 2)An additional option for training named "loss-weight" to change the aforementioned branch loss weight. 3)Some modifications in yolo.py and common.py to adapt to the new variant.