Temporal video analysis

[MalteEbner]

Does MMPose support a temporal analysis of videos yet, where the pose estimation is not done on a frame-by-frame basis, but rather using temporal information, like in this paper on VideoPose3D ?
Supporting temporal frameworks would need the datasets to consist of an unordered set of videos, where each video is a sequential ordered list of frames. Furthermore, the splitting of training and validation set, the definition of batches and the input to the model would also need to work accordingly.

[jin-s13]

Currently, MMPose does not support video pose estimation yet.
Thank you for the suggestion. I will leave a comment in #9 and we may consider adding this feature in the following versions.

[F9393]

Hello,

I am wondering if the mmpose repo currently supports video pose estimation (by taking into account temporal information)?

Thanks

[innerlee]

Hi @FarzanehAskari , for now supported video methods, checkout https://mmpose.readthedocs.io/en/latest/topics/body(3d,kpt,vid).html

We will add more in future, follow our monthly iteration plan for news!

[F9393]

Hello, I am wondering where to find the update for which video methods are included as of now? Thank you for the great repo.

[ly015]

We currently support 3D video method (checkout here) VideoPose3D (CVPR'19) and 2D video method (checkout here) PoseWarper (NeurIPS'19). And we are also working on post-processing methods to improve temporal smoothness and robustness (checkout #1127).

Please follow our monthly iteration plan and release notes for updates.