Abstract
We introduce a new approach for 2D human pose estimation in stereo video sequences. Our proposed pipeline starts by constraining the possible location of body joints by exploding color and disparity information, and adding location priors to the shoulders. Then, a body limbs recombination technique is applied along the stereo sequence to obtain the best configuration of the body joints.
We validate our model on three challenging datasets: Stereo Human Pose Estimation Dataset that contains stereo video sequences, Poses in the Wild that contains monocular video sequences, and Inria 3DMovie Dataset that contains stereo image pairs.
The experimental results indicates that our model establish new state-of-the-art results on stereo sequences and improves on monocular models.
Supplementary Information
- Best Limb Score graphical model [PDF].
- Comparison of accuracy on SHPED (max.) dataset [PDF].
- Comparison of accuracy on PIW (avg.) dataset [PDF].
- Comparison of accuracy on PIW (max.) dataset [PDF].
Paper
This work has been published as:
Mixing body-parts model for 2D human pose estimation in stereo videos
IET Computer Vision, vol. 11, issue 6, pp. 426-433, 2017
Acknowledgements
This work was partially supported by the Research Projects TIN2012-32952 and BROCA, both financed by the Spanish Ministry of Science and Technology and the European Regional Development Fund (FEDER).