Detail of training

[neo85824]

Hi, I'm trying to train yolact on cityscapes dataset with my code. However, my result just looks worst.
I've traced your cityscapes training config. Did you only modify the image size, learning rate and lr_steps on training compared to the original yolact config?
or there is some additional modification or trick? Thank you.

[KevinJia1212]

The flag USE_SQUARE_ANCHORS in the original yolact_v1 config is set to True, this makes it hard for the network to detect objects with large aspect ratios like pedestrains.