Haodong Duan

Hi, ahmed-nady, That's a common practice in general image / video recognition. For example, many works using 224 x 224 as the training resolution, and 256 x 256 as the...

Emmm, please check the forward function. Actually, we are just doing what u described.

For the later choice, you will lose the semantic information about each point: how can you tell the difference of two joints if they are just two points in the...

For visualization, you can use https://github.com/kennymckormick/pyskl/blob/main/demo/visualize_heatmap_volume.ipynb, it converts keypoint coordinates to 3D heatmap volumes and visualize them. If you want to do it in training, that is implemented as data...

Hi, @sasakits , would you please list the platform that you run this command (for example, the torch version, cuda version, and the GPU model).

It seems your platform only has 1 GPU, so you should use `!bash tools/dist_run.sh tools/data/custom_2d_skeleton.py 1 (# This is the GPU number) --video-list examples/extract_diving48_skeleton/diving48.list --out examples/` instead.

Then I'm not sure about the problem, maybe you can try with an early version of mmcv, like mmcv==1.5.0

Hi, sasakits, recently I have supported an alternative skeleton extraction method, which runs in non-distributed mode. You can check the ipython notebook and try. Also, you can follow the [new...

In the `config` directory, all configs with tag `ntu` and `3dkp` in its path are configs of NTURGB+D(3D skeleton)

1. PoseConv3D uses 3D conv for temporal modeling, if replacing it with normal 2D conv, the temporal modeling capability may significantly deteriorate and the recognition accuracy will drop. 2. It's...