PoseNet

Distilling 3D Human Pose from a Single Image

---

Code & Report for course Machine Perception at ETH Zurich, Spring 2019.

Reproducing Results

Install Requirements

make requirements
mkdir log/

Note: All configuration files provided assume that all available data is placed at /cluster/project/infk/hilliges/lectures/mp19/project2/

Training from scratch

1. Training only HRNet
   • trains only for 2D pose estimation.
   • applies augmentation (horizontal flip).
   • trained for 1 epoch.
   • makes use of config/pretrain.py
   • weights stored at log/pretrain/
   • if configuration unchanged, the name should be PRETRAIN-master-Adam-1
   • should take around 4 hours.

make submit SCRIPT=scripts/pretrainHRN.py Q=6:00

2. Train entire model using pretrained-HRNet
   • uses weights from step 1 as initialization for HRNet.
   • makes use of config/posenet.py
   • configuration loads weights directly from step 1. In case of any issues, kindly set the variable PRETRAINED on line 41 of config/pretrain.py to the path of the step 1 weights.
   • trained for 8 epochs.
   • produces submission files after every epoch. The last one is the submitted to the leaderboard.
   • should take around additional 30 hours.

make submit SCRIPT=main.py Q=32:00

Pretrained weights

For sake of convienience, I have uploaded the pretrained weights of step 1 to polybox. One can download and place them at log/pretrain/ and start directly with step 2.

wget https://polybox.ethz.ch/index.php/s/YwitdaXVXN31UWB/download -O log/pretrain/PRETRAIN-master-Adam-1

Run-time logs

Logs are flushed to file, which can be found in log/pretrain/ for pretraining and log/master/ for step 2. Users can also stream the log outputs using the command make stream DIR=<pretrain/master> Logs are saved in the format {dd}-{mm}--{H}-{s} in respective directories.