by Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, Jiaya Jia, details are in project page.
This repository is modified from Caffe version of yjxiong and DeepLab v2 for testing. Results are tested on Ubuntu 14.04. Trainable code will be available later.
-
Clone the repository:
git clone https://github.com/hszhao/PSPNet.git
-
Build Caffe and matcaffe
cd $PSPNET_ROOT make -j8 && make matcaffe
-
Testing
Evaluation code is in folder 'evaluation'.
Download trained models and put it in folder 'evaluation/model/':
pspnet50_ADE20K.caffemodel: link
pspnet101_VOC2012.caffemodel: link
pspnet101_cityscapes.caffemodel: link
Modify the related paths in 'eval_all.m':
cd evaluation vim eval_all.m
Run the testing scripts:
./run.sh
-
Results: (single scale testing denotes as 'ss' and multiple scale testing denotes as 'ms')
- PSPNet50 on ADE20K valset (mIoU/pAcc): 41.68/80.04 (ss) and 42.78/80.76 (ms)
- PSPNet101 on VOC2012 testset (mIoU): 85.41 (ms)
- PSPNet101 on cityscapes valset (mIoU/pAcc): 79.70/96.38 (ss) and 80.91/96.59 (ms)
Please contact '[email protected]'