search

hellochick / ICNet-tensorflow

TensorFlow-based implementation of "ICNet for Real-Time Semantic Segmentation on High-Resolution Images".
406 stars 153 forks source link

The result(mIoU) of cityscapes is too bad. #54

Open L6011 opened 6 years ago

L6011 commented 6 years ago

I ran the evaluation.py using the model of icnet_cityscapes_train_30k_bnnomerge.npy directly, but the mIoU was only 13.8%. Then I used the icnet_cityscapes_train_30k_bnnomerge.npy as a initial model and trained my own model, but the result still 13%. My bachsize was 8, other parameters just as yours, and iterated 3000 times. Do you have any solution?

hellochick commented 6 years ago

Maybe you haven't converted your annotations yet? Please see the tutorial follow the README, you should use the provided cityscapes script first.