PSPNet implemented in PyTorch for single-person human parsing task, evaluating on Look Into Person (LIP) dataset.
The implementation of PSPNet is based on HERE.
Trained model weights can be downloaded from Google Drive or Baidu Drive (提取码：43cu).
- Python 3.6
- PyTorch == 1.1.0
- torchvision == 0.3.0
- matplotlib
To use our code, firstly you should download LIP dataset from HERE.
Then, reorganize the dataset folder as below:
myLIP
│
└───train
│ │ id.txt
│ │
│ └───image
│ │ │ 77_471474.jpg
│ │ │ 113_1207747.jpg
│ │ │ ...
│ │
│ └───gt
│ │ │ 77_471474.png
│ │ │ 113_1207747.png
│ │ │ ...
│
└───val
│ │ id.txt
│ │
│ └───image
│ │ │ 100034_483681.jpg
│ │ │ 10005_205677.jpg
│ │ │ ...
│ │
│ └───gt
│ │ │ 100034_483681.png
│ │ │ 10005_205677.png
│ │ │ ...
│
└───test
│ │ id.txt
│ │
│ └───image
│ │ │ 100012_501646.jpg
│ │ │ ...
python3 train.py --data-path PATH-TO-LIP --backend [resnet50 | densenet | squeezenet]
python3 eval.py --data-path PATH-TO-LIP --backend [resnet50 | densenet | squeezenet] [--visualize]
python3 inference.py demo/test.jpg --backend [resnet50 | densenet | squeezenet]
| model | overall acc. | mean acc. | mean IoU |
|---|---|---|---|
| resnet50 | 0.792 | 0.552 | 0.463 |
| resnet101 | 0.805 | 0.579 | 0.489 |
| densenet121 | 0.826 | 0.606 | 0.519 |
| squeezenet | 0.786 | 0.543 | 0.450 |
>> python3 eval.py --data-path PATH-TO-LIP --visualize
>> python3 inference.py demo/test.jpg
