Repository for the Sports Shape and Pose 3D (SSP-3D) dataset, as introduced in "Synthetic Training for Accurate 3D Human Pose and Shape Estimation in the Wild".
SSP-3D is an evaluation dataset consisting of 311 images of sportspersons in tight-fitted clothes, with a variety of body shapes and poses. The images were collected from the Sports-1M dataset. SSP-3D is intended for use as a benchmark for body shape prediction methods. Pseudo-ground-truth 3D shape labels (using the SMPL body model) were obtained via multi-frame optimisation with shape consistency between frames (as described in the paper above). The figure above shows a few samples from the SSP-3D dataset (under 'Optimised Pose and Shape') as well as a comparison with pre-optimised body predictions (using VIBE), which demonstrates the improvement in body model fit achieved using optimisation.
Since SSP-3D is a small dataset, the zip file containing all the necessary data is a part of this repository. Unzipping it will reveal a folder with images, a folder with silhouette annotations and a file called labels.npz. This file contains arrays with filenames, SMPL pose parameters, SMPL shape parameters, genders, 2D joint annotations, camera translations and bounding boxes for each image.
We provide a python3 script visualisation.py that demonstrates how to project/render/visualise all the relevant data. To run the script, you will need to install the relevant libraries: pip install requirements.txt. If you have trouble install pyrender, please take a look at the docs.
Additionally, please download the SMPL male and female models. You will need to convert the SMPL model files to be compatible with python3 by removing any chumpy objects. To do so, please follow the instructions here. Rename the models to SMPL_MALE.pkl and SMPL_FEMALE.pkl. Finally, set SSP_3D_PATH in config.py to the SSP-3D root directory path. Set SMPL_MODEL_DIR to the path of the directory with the SMPL models.
After completing set-up, run python visualisation.py.
This dataset is intended for use as a body shape evaluation benchmark. If you decide to use this dataset, we recommend you report the following metrics:
- mIOU: mean Intersection-over-Union between target silhouette and predicted SMPL silhouette.
- PVE-T-SC (mm): per-vertex error in a neutral pose (T-pose) after after scale-correction (to account for scale vs camera depth ambiguity). An example of how to compute this metric given predicted in
metrics.py.
If you find this dataset useful for your research, please cite the following publication:
@InProceedings{STRAPS2018BMVC,
author = {Sengupta, Akash and Budvytis, Ignas and Cipolla, Roberto},
title = {Synthetic Training for Accurate 3D Human Pose and Shape Estimation in the Wild},
booktitle = {British Machine Vision Conference (BMVC)},
month = {September},
year = {2020}
}