Maysee
Maysee
I'm not sure if we can convert 17 joints to 25 joints. But you may train a model to output 25 joints. This issue may help https://github.com/facebookresearch/VideoPose3D/issues/151. Sorry for the...
现在用的2D to 3D 模型[VedioPose3D](https://github.com/facebookresearch/VideoPose3D)无法做到实时,可以考虑把VideoPose3D换成更轻量级的模型。
Check #48.
3D输出的单位是米,但是因为输入的2D坐标被归一化过了,因此这只是给3D空间作图用的单位,不能反映真是世界大小。
Sorry that I have no idea how to achieve this Moving-Grid rendering.
I'm working on my paper recently. I may try to add it next month.
I don't train the model. The checkpoint is definitely from the official model: https://github.com/facebookresearch/VideoPose3D/blob/master/INFERENCE.md#step-1-setup.
Check https://github.com/MVIG-SJTU/AlphaPose/issues/557
This link may help: https://stackoverflow.com/questions/62213783/ffmpeg-error-while-loading-shared-libraries-libopenh264-so-5. I'm currently using `ffmpeg version 3.4.8-0ubuntu0.2 Copyright (c) 2000-2020 the FFmpeg developers`
1. The definition of rot: check this [issue](https://github.com/zh-plus/video-to-pose3D/issues/4#issuecomment-561252908). 2. This rotation value is retrieved from a random object. In the original script from Facebook, there's a comment indicating that: [Comments](https://github.com/zh-plus/video-to-pose3D/blob/e271582b762f9f56b50f75f8d44cc351736d8683/origin_run.py#L737)...