search

yu-changqian / TorchSeg

Fast, modular reference implementation and easy training of Semantic Segmentation algorithms in PyTorch.
MIT License
1.41k stars 253 forks source link

evaluation problem？ #78

Open guyuchao opened 5 years ago

guyuchao commented 5 years ago

I find you do resize on the label to compute iou instead of resize your pred to origin label size? This may be iccurrate in iou because you loss details in small object.
label = cv2.resize(label, (config.image_width // config.gt_down_sampling, config.image_height // config.gt_down_sampling), interpolation=cv2.INTER_NEAREST)

yu-changqian commented 5 years ago

Yes. You can also change the evaluation code to the normal size with the same change to the training code.