Implementing Pyramid Scene Parsing Network (PSPNet) paper using Pytorch
- PSPNet
- Reference
- Paper
- Author: Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, Jiaya Jia
- Organization: The Chinese University of Hong Kong, SenseTime Group Limited
- Download VOC2012 Dataset
- Data Tree
Dataset
├── prepare_voc.py
├── VOC2012
│ ├── Annotations
| ├── Images
│ ├── ImageSets
│ │ | ├── Action
│ │ | ├── Layout
│ │ | ├── Main
│ │ | ├── Segmentation
│ │ | | ├── train.txt
│ │ | | ├── trainaug.txt
│ │ | | ├── trainaugval.txt
│ │ | | ├── trainval.txt
│ │ | | ├── val.txt
│ ├── Labels
| ├── SegmentationClass
│ ├── SegmentationObject
- Prepare VOC2012
python dataset/prepared_voc.py
- Train
- 1 GPU
python main.py --evaluation False - Multi GPUs (ex, 8 GPUs)
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --nproc_per_node=8 --master_port=$RANDOM main.py
- Test
python main.py --evaluation True
| Datasets | Resource | Model | Mean IoU | Pixel Acc |
|---|---|---|---|---|
| VOC2012 | 8 GPUs | PSPNet50 (Baseline) | 0.7802 | 0.9513 |
| VOC2012 | 8 GPUs | PSPNet50 (Our) | 0.7644 | 0.9400 |
| VOC2012 | 8 GPUs | PSPNet50 (Our + Skip(Out1)) | 0.77414 | 0.9431 |
| VOC2012 | 8 GPUs | PSPNet50 (Our + Skip(Out1, Stem)) | 0.78141 | 0.9501 |
| VOC2012 | 1 GPU | PSPNet50 (Our) | 0.7378 | 0.9322 |
| VOC2012 | 8 GPUs | PSPNet101 (Baseline) | 0.7963 | 0.9550 |
| VOC2012 | 8 GPUs | PSPNet101 (Our) | 0.7931 | 0.9487 |