The result finetune with cityscapes is very poor #141
Description
The COCO-trained model is used as a pre-training model to finetune the cityscapes dataset. The test results are poor and the edges overlap.
I don't think it should be like this. Has anyone done this? How's the effect?