Awesome

Transforming Self-Supervision in Test Time for Personalizing Human Pose Estimation

This is an official implementation of the NeurIPS 2021 paper: Transforming Self-Supervision in Test Time for Personalizing Human Pose Estimation. More details can be found at our project website.

Preparation

1. Install dependencies

pip install -r requirements.txt

2. Make libs

   cd ${PROJECT_ROOT}/lib
   make
   

3. Place Penn Action data in data directory. (Instructions on Human3.6M and BBC Pose are coming soon.)

   Your directory tree should look like this:

   ${PROJECT_ROOT}
   └── data
       └── Penn_Action
           ├── frames
           ├── labels
           ├── tools
           └── README
   

4. Download pretrained model of ResNet-18 and ResNet-50 and place them in models/pytorch/imagenet.

   Your directory tree should look like this:

   ${PROJECT_ROOT}
   └── models
       └── pytorch
           └── imagenet
               ├── resnet18-5c106cde.pth
               └── resnet50-19c8e357.pth
   

Training and Test-time Personalization

Training

python tools/train_joint.py \
   --cfg experiments/penn/joint_res50_128x128_1e-3_comb_attn_tf1_4head.yaml

Run Test-Time Personalization (online)

python tools/test_time_training.py \
   --cfg experiments/penn/ttp_res50_128x128_lr1e-4_online_downsample1_comb_attn_tf1_4head.yaml \
   TEST.MODEL_FILE ${MODEL_FILE}

Run Test-Time Personalization (offline)

python tools/test_time_training.py \
   --cfg experiments/penn/ttp_res50_128x128_lr1e-4_offline_downsample1_comb_attn_tf1_4head.yaml \
   TEST.MODEL_FILE ${MODEL_FILE}

Baseline Model

To train the baseline model for comparison

python tools/train.py --cfg experiments/penn/res50_128x128.yaml

Result

Configs, results and model checkpoints on Human3.6M and BBC Pose are coming soon.

Acknowlegement

TTP is developed based on HRNet. We also incorperate some code from IMM.