search

meetps / pytorch-semseg

Semantic Segmentation Architectures Implemented in PyTorch
https://meetshah.dev/semantic-segmentation/deep-learning/pytorch/visdom/2017/06/01/semantic-segmentation-over-the-years.html
MIT License
3.38k stars 800 forks source link

fcn8s on Pascal VOC #120

Open hihfs opened 5 years ago

hihfs commented 5 years ago

Hi,

I am trying to reproduce the results in the PAMI version of the FCN paper, training fcn8s on Pascal VOC with the default setting. In the paper (Table 4), the authors report the mean IOU of 67.2 on VOC2012. I can't seem to achieve anything near it on the validation set during the training (currently it is ~62). What is the difference, do you suppose?

Also, I am wondering if the argument 'split' in the code:

    v_loader = data_loader(
        data_path,
        is_transform=True,
        split="val",
        img_size=(args.img_rows, args.img_cols),
        img_norm=args.img_norm,
    )

in train.py should be "train_aug_val" instead of "val". As it is, the validation set has an intersection with the training set "train_aug". If we use it, wouldn't the IOU become even lower? Also, which set in your repo is the "test set" authors of the above paper refer to?

Feiyu-Zhang commented 5 years ago

Mean IOU is 65.7 on the validation set. But I don's test on the test set.