BiHMP-GAN: Bidirectional 3D Human Motion Prediction GAN #11
Description
1. Overview (basic ideas)
Framework for predicting human pose using wGANs
2. Novelty
While other methods train a recurrent neural network directly, this paper first learn s a continuous pose embedding space independent of motion dynamics. This helps avoid prediction of unlikely or improbable skeleton joint parameters and is especially effective for learning long term (200ms~) motion.
3. Method (Technical details)
4. Results
5. links to papers, codes, etc.
arxiv: https://arxiv.org/abs/1812.02591
6. Thoughts, Comments
Maybe incorporating the Grassman manifold to learn the pose manifold might be interesting?
The discriminator architecture is a simple idea and I think it won't work on longer poses. Its too difficult to train GANs this naively.
A reinforcement learning setup might help too? I wonder if someone has already worked on a RL loss for future poses. Maybe the engineering overhead is way too large tho.
7. bibtex
@misc{kundu2018bihmpgan,
title={BiHMP-GAN: Bidirectional 3D Human Motion Prediction GAN},
author={Jogendra Nath Kundu and Maharshi Gor and R. Venkatesh Babu},
year={2018},
eprint={1812.02591},
archivePrefix={arXiv},
primaryClass={cs.CV}
}