search

sacmehta / ESPNet

ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation
https://sacmehta.github.io/ESPNet/
MIT License
541 stars 112 forks source link

The IoU calculation problem. #33

Closed zhengxiawu closed 5 years ago

zhengxiawu commented 5 years ago

in line 41 of train/IOUEval.py

self.batchCount += 1

but in train/main.py, the batch size is batch_size=args.batch_size + 4

when I set batch_size = 1, the performance of decoder_2_8 on cityscape will drop to 37.05 and on camVid will drop to 44.89。 on camVid when batch size = 1 mIOU = 44.89. batch size = 4 mIOU=45.9, batch size = 32 mIOU = 44.4

sacmehta commented 5 years ago

IOUEval should be independent of the provided batch size (via args) because it computes the batch size based on the input and target matrices. Let me take a more closer look.

zhengxiawu commented 5 years ago

thanksa lot! (the provided network works fine with cityscape scripts, but lower when using IOUEval)would you please tell me how u evaluate the performance on camVid? on ioueval ?

zhengxiawu commented 5 years ago

I also have done more experiment on cityscape with ESPnet, i.e. batch_size=1, mIOU=37.05; batch_size=2,mIOU=40.7;batch_size=4,mIOU=43.9;

sacmehta commented 5 years ago

Could you share your procedure for computing mIOU at different batch sizes?