Could you provide some example on how to use the hand pose estimator about trihorn-net

[mrezaei92]

Dear @TheoCarme,

Please note that TriHorn-Net takes a single image as input. If you plan on performing hand pose estimation on a video stream, the network architecture should be modified. To this end, you need to consider some factors such as:

• Are you going to predict hand pose for the latest frame in the video, or are you going to perform hand pose estimation on each frame in the sequence by taking advantage of temporal information embedded in the video?
• What dataset are you going to use? If you plan on working on RGB images, there are some video datasets, but in the realm of depth images, as far as I know, there is only NYU. You should note that there are some discontinuities in the frames of NYU, which have to be detected manually.

[TheoCarme]

Dear @mrezaei92 thank you for responding.

When I wrote "on a video stream" I meant : to perform the hand pose estimation one frame at the time as I grab them.
As for the dataset I intend to use MSRA since contrary to ICVL and NYU it has 21 joints per frame.
So could you provide me with some examples please ?

Also I was wondering if it is normal that when I trained on MSRA, at the end of the 60th epoch it automatically restarted the training at the first epoch.

[TheoCarme]

Dear @mrezaei92 could you please help me ?