zbf1991 / RRM

Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach

Reliability Does Matter: An End-to-End Weakly Supervised Semantic Segmentation Approach

AAAI 2020 (Spotlight).

The code of the new extended work now is available. In the further, I will try to merge these two works so that making the whole project is more elegant.

Paper of the extended work will come soon.

This project is based on Regularized loss and PSA.

Before Running, build python extension module:

cd wrapper/bilateralfilter
swig -python -c++ bilateralfilter.i
python setup.py install

More details please see here

Download pretrained models to ./netWeights:

Google: due to the coronavirus outbreak in China, I will upload models after I can enter my lab. But you can download “[ilsvrc-cls_rna-a1_cls1000_ep-0001.params]” and “[res38_cls.pth]” from here.

BaiduYun

[ilsvrc-cls_rna-a1_cls1000_ep-0001.params] is an init pretained model.

[res38_cls.pth] is a classification model pretrained on VOC 2012 dataset.

[RRM_final.pth] is the final model (AAAI).

[RRM(attention)_final.pth] is the final model of the new extended work (64.7 mIoU on Pascal Voc 2012 val set).

Training of the New Extended Work:

Training from init model:

you need 4 GPUs and the pretrained model [ilsvrc-cls_rna-a1_cls1000_ep-0001.params]:

python train_from_init(attention).py --voc12_root /your/path/VOCdevkit/VOC2012

Training from a pretrained classification model:

you only need 2 GPU and the pretrained model [res38_cls.pth]

python train_from_cls_weight(attention).py --IMpath /your/path/VOCdevkit/VOC2012/JPEGImages

Training of the AAAI Work:

I suggest that it is better to use the 2nd method due to lower computing costs.

Training from init model:

you need 4 GPUs and the pretrained model [ilsvrc-cls_rna-a1_cls1000_ep-0001.params]:

python train_from_init.py --voc12_root /your/path/VOCdevkit/VOC2012

Training from a pretrained classification model:

you only need 1 GPU and the pretrained model [res38_cls.pth]

python train_from_cls_weight.py --IMpath /your/path/VOCdevkit/VOC2012/JPEGImages

Inferencing of the extended work:

you need 1 GPU and the final model [RRM(attention)_final.pth]:

python infer_RRM.py --IMpath /your/path/VOCdevkit/VOC2012/JPEGImages

Inferencing of the AAAI work:

you need 1 GPU and the final model [RRM_final.pth]:

python infer_RRM.py --IMpath /your/path/VOCdevkit/VOC2012/JPEGImages