1. Clone this repo, and we'll call the directory that you cloned multiview-pose as ${POSE_ROOT}, and enter the path of single-view 2D pose estimation

git clone https://github.com/HowieMa/PPT.git
cd single-view-PPT/

2. Install dependencies.

pip install -r requirements.txt

3. Download pretrained models from our model zoo: (Google Drive). We also put the pre-trained HRNet, TransPose, and TokenPose (rerun by us) into the same folder to simplify the downloading. All copyright belongs to their respective owners! You can also download them from their original source separately:
   HRNet model zoo
   TransPose model zoo

Please put them under ${POSE_ROOT}/single-view-PPT/models, and make them look like this:

${POSE_ROOT}/single-view-PPT/models
└── pytorch
    └── imagenet
        └── hrnet_w32-36af842e.pth
        └── hrnet_w48-8ef0771d.pth
        └── ...
    └── pose_coco
        └── ppt_s_256x192_ratio07.pth
        └── ppt_b_256x192_ratio07.pth
        └── ppt_l_d6_256x192_ratio07.pth
        └── ... 

4. Data Preparation We follow the steps of HRNet to prepare the COCO train/val/test dataset and the annotations. The detected person results are downloaded from GoogleDrive. Please download or link them to ${POSE_ROOT}/data/coco/, and make them look like this

${POSE_ROOT}/single-view-PPT/data/coco/
|-- annotations
|   |-- person_keypoints_train2017.json
|   `-- person_keypoints_val2017.json
|-- person_detection_results
|   |-- COCO_val2017_detections_AP_H_56_person.json
|   `-- COCO_test-dev2017_detections_AP_H_609_person.json
`-- images
	|-- train2017
	|   |-- 000000000009.jpg
	|   |-- ... 
	`-- val2017
		|-- 000000000139.jpg
		|-- ... 

We take PPT-S as an example:

1. Testing on COCO val2017 dataset

python tools/test.py --cfg experiments/coco/ppt/ppt_s_v1_256_192_patch43_dim192_depth12_heads8.yaml TEST.MODEL_FILE models/pytorch/pose_coco/ppt_s_256x192_ratio07.pth TEST.USE_GT_BBOX False

This script should create folders {POSE_ROOT}/single-view-PPT/output/coco/pose_tokenpose_s/ppt_s_v1_256_192_patch43_dim192_depth12_heads8/
The result should be

2. Compute FPS and Throughput

python tools/compute_fps.py --cfg experiments/coco/ppt/ppt_s_v1_256_192_patch43_dim192_depth12_heads8.yaml

On a single 2080Ti, the FPS should be 130 and the throughput should be 870 images/s Note that, the results may be slightly different due to randomness.

3. Training on COCO train2017 dataset

python tools/train.py --cfg experiments/coco/ppt/ppt_s_v1_256_192_patch43_dim192_depth12_heads8.yaml

For all training / testing scripts, please see scripts.sh