/3d-pose-baseline

A simple baseline for 3d human pose estimation in tensorflow. Presented at ICCV 17.

Primary LanguagePythonMIT LicenseMIT

3d-pose-baseline

This is the code for the paper

Julieta Martinez, Rayat Hossain, Javier Romero, James J. Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. https://arxiv.org/pdf/1705.03098.pdf.

The code in this repository was mostly written by Julieta Martinez, Rayat Hossain and Javier Romero.

We provide a strong baseline for 3d human pose estimation that also sheds light on the challenges of current approaches. Our model is lightweight and we strive to make our code transparent, compact, and easy-to-understand.

Dependencies

First of all

  1. Watch our video: https://youtu.be/Hmi3Pd9x1BE
  2. Clone this repository and get the data.

[Update May 22, 2020] We are in the process of putting together a tutorial to prepare the data for this code. Please check back soon for an update.

git clone https://github.com/una-dinosauria/3d-pose-baseline.git
cd 3d-pose-baseline

Quick demo

For a quick demo, you can train for one epoch and visualize the results. To train, run

python src/predict_3dpose.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh --epochs 1

This should take about <5 minutes to complete on a GTX 1080, and give you around 75 mm of error on the test set.

Now, to visualize the results, simply run

python src/predict_3dpose.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh --epochs 1 --sample --load 24371

This will produce a visualization similar to this:

Visualization example

Training

To train a model with clean 2d detections, run:

python src/predict_3dpose.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise

This corresponds to Table 2, bottom row. Ours (GT detections) (MA)

To train on Stacked Hourglass detections, run

python src/predict_3dpose.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh

This corresponds to Table 2, next-to-last row. Ours (SH detections) (MA)

On a GTX 1080 GPU, this takes <8 ms for forward+backward computation, and <6 ms for forward-only computation per batch of 64.

Pre-trained model

We also provide a model pre-trained on Stacked-Hourglass detections, available through google drive.

To test the model, decompress the file at the top level of this project, and call

python src/predict_3dpose.py --camera_frame --residual --batch_norm --dropout 0.5 --max_norm --evaluateActionWise --use_sh --epochs 200 --sample --load 4874200

Fine-tuned stacked-hourglass detections

You can find the detections produced by Stacked Hourglass after fine-tuning on the H3.6M dataset on google drive.

Citing

If you use our code, please cite our work

@inproceedings{martinez_2017_3dbaseline,
  title={A simple yet effective baseline for 3d human pose estimation},
  author={Martinez, Julieta and Hossain, Rayat and Romero, Javier and Little, James J.},
  booktitle={ICCV},
  year={2017}
}

Other implementations

Extensions

License

MIT