- Pytorch official code of QANet is released. (03/11/2021)
Official implementation of Quality-Aware Network for Human Parsing
In this repository, we release the QANet code in Pytorch.
- QANet architecture:
If you use QANet, please use the following BibTeX entry.
@inproceedings{yang2021qanet,
title = {Quality-Aware Network for Human Parsing},
author = {Lu Yang and Qing Song and Zhihui Wang and Zhiwei Liu and Songcen Xu and Zhihao Li},
booktitle = {arXiv preprint arXiv:2103.05997},
year = {2021}
}
- 8 x TITAN RTX GPU
- pytorch1.6
- python3.6.8
Install QANet following INSTALL.md.
Please follow DATA_PREPARE.md to download training and evaluating data.
QANet On CIHP
| Backbone | mIoU | APp/APp50/PCP50 | APr/APr50 | DOWNLOAD |
|---|---|---|---|---|
| ResNet50 | 62.9 | 60.1/74.3/68.9 | 56.2/63.5 | GoogleDrive |
| ResNet101 | 64.1 | 62.0/77.9/72.4 | 57.9/65.6 | |
| HRNet-W48 | 66.1 | 64.5/81.3/75.7 | 60.8/68.8 | GoogleDrive |
QANet On LIP
| Backbone | Input Size | pixAcc. | meanAcc. | mIoU | DOWNLOAD |
|---|---|---|---|---|---|
| HRNet-W48 | 512×384 | 88.92 | 71.87 | 59.61 | GoogleDrive |
| HRNet-W48 | 544×416 | 89.19 | 72.97 | 60.52 | GoogleDrive |
- Flip test is used.
- For CIHP, we use FCOS-R50 to detect person (73.1 AP on CIHP val).
- Multi-scale test is used for LIP.
ImageNet pretrained weights
please put the pretrained weights in QANet/weights
To train a model with 8 GPUs run:
python tools/train_net_all.py --cfg cfgs/CIHP/QANet/QANet_R-50c_512x384_1x.yaml
python tools/test_net_all.py --cfg ckpts/CIHP/QANet/QANet_R-50c_512x384_1x/QANet_R-50c_512x384_1x.yaml --gpu_id 0,1,2,3,4,5,6,7
python tools/test_net_all.py --cfg ckpts/CIHP/QANet/QANet_R-50c_512x384_1x/QANet_R-50c_512x384_1x.yaml --gpu_id 0
QANet is released under the MIT license.