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

AAAI 2020 (Spotlight).

Due to the coronavirus outbreak in China, I cannot return to my lab, this project is uploaded throuthing the remote desktop. I will rewrite this README file and answer issuses after I can go back.

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 my final model. mIoU is about 63.7 on val set, which is a higher score than our paper (62.6)

Training:

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 --IM_path /your/path/VOCdevkit/VOC2012/JPEGImages

Inferencing:

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

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