The result finetune with cityscapes is very poor

msracver / FCIS

Fully Convolutional Instance-aware Semantic Segmentation

MIT License

1.57k stars 415 forks source link

The result finetune with cityscapes is very poor #141

Open vtHuo opened 5 years ago

vtHuo commented 5 years ago

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?