Seeing the Unseen:
Wifi-based 2D Human Pose Estimation via Dilated Evolving Attentive Spatial-Frequency Network
| GPE | SPE | ||||
| # Person | Number | Url | Action | Number | Url |
| 1 | 12,884 | https://reurl.cc/EzXD8R | Walk | 78,852 | https://reurl.cc/OqEL9D |
| 2 | 18,879 | https://reurl.cc/ygmX1a | Wave | 77,431 | https://reurl.cc/Z71vGQ |
| 3 | 27,694 | https://reurl.cc/e8Wa5W | Jump | 40,670 | https://reurl.cc/j5RYNM |
| >=4 | 28,178 | https://reurl.cc/m9ZXd7 | Run | 41,238 | https://reurl.cc/EzXL3m |
| Cross Period | X | X | Cross Period | X | https://reurl.cc/ld4mZq |
| Total | 87,635 | X | Total | 238,191 | X |
$ git clone https://github.com/DEASFN/DEASFN.git
- create the virtual enciorment
$ cd DEASFN/
$ python3 -m venv env
- activate the virtual environment
$ source env/bin/activate
- install the necessary packages
$ pip install -r requirements.txt
If there are errors about gdown package, you can ignore them.
- Download SPE or GPE dataset
$ chmod +x download_SPE.sh or $ chmod +x download_GPE.sh
$ ./download_SPE.sh or $ ./download_GPE.sh
When you find the download has exited, you need to press ctrl+c to terminate the sh file by yourself.
If errors occur, you can download by yourself using the links provided above.
Please make sure that all four zipfiles in DEASFN/dataset/SPE or DEASFN/dataset/GPE.
- unzip the dataset ( You can substitute
GPEforSPE)
$ cd DEASFN/dataset/SPE
$ chmod +x SPE.sh
$ ./SPE.sh
You can adjust the training parameters in config/args.py
$ python3 TrainDeasfn.py--dataset=SPE or $ python3 TrainDeasfn.py --dataset=GPE
You can adjust the testing parameters in config/args.py
$ python3 TestDeasfn.py --dataset=SPE or $ python3 TestDeasfn.py --dataset=GPE
| DEASFN | PIW | WiSPPN | |
| Walk |  |
 |
 |
| Wave |  |
 |
 |
| Jump |  |
 |
 |
| Run |  |
 |
 |
From this comparison, we can easily observe that DEASFN is better than the camera-based method in dark environment. Under poor illumination, Openpose is likely to loss some keypoints, such as wrist, ankle, and so forth. However, DEASFN can correctly return all keypoints.
| DEASFN | Openpose |
|---|---|
 |
 |
PCK curve of each model:
PCK of DEASFN on each joint:
- Comparisons on the proposed benchmarks:
| Metric | Benchmark | WiSPPN[1] | PIW[2] | DEASFN |
| MPJPE | SPE | 44.16 | 78.88 | 37.34 |
| GPE | X | 119.60 | 44.14 | |
| PCK@20 | SPE | 21.86% | 32.96% | 50.05% |
| GPE | X | 27.64% | 43.98% |
- Comparative results against other methods on our SPE benchmark (PCK@20):
| Action | WiSPPN | PIW | DEASFN |
|---|---|---|---|
| Walk | 23.58% | 39.87% | 61.14% |
| Wave | 25.92% | 33.06% | 45.81% |
| Run | 22.12% | 37.91% | 58.11% |
| Jump | 15.82% | 20.99% | 35.15% |
- Comparative results against other methods on our GPE benchmark (PCK@20):
| # Person | PIW | DEASFN |
|---|---|---|
| 1-person | 53.37% | 72.65% |
| 2-person | 49.31% | 65.39% |
| 3-person | 19.54% | 34.75% |
| >=4 person | 22.66% | 39.12% |
- Ablation study on our SPE benchmark:
| Model | Architecture | PCK@20 |
|---|---|---|
| 1 | SFN | 44.99% |
| 2 | ASFN | 45.95% |
| 3 | DEASFN, D=1 | 48.08% |
| 4 | DEASFN, D=3 (Proposed) | 50.05% |
| 5 | DEASFN, D=5 | 45.95% |
- [1] Fei Wang, Stanislav Panev, Ziyi Dai, Jinsong Han, and Dong Huang. 2019. Can wifi estimate person pose?arXiv preprint arXiv:1904.00277(2019).
- [2] Fei Wang, Sanping Zhou, Stanislav Panev, Jinsong Han, and Dong Huang. 2019.Person-in-WiFi: Fine-grained person perception using WiFi. InProceedings of theIEEE International Conference on Computer Vision. 5452–5461.