About pose tracking

[XiangpengZeng]

Hello！Sorry to interrupt you. I have also paid attention to this article recently. Thank you for your work. However, some parts of the article confuse me. I would like to consult that what exactly is pose track and what is the difference between it and general pose estimation（such as openpose）? , Thank you very much for your answers

[jiang-du]

Good question. In my point of view, the difference between pose tracking and pose estimation is as follows.

1. The task.

• In pose estimation task, we only have an input RGB image and nothing more. The task is just to infer the skeleton joints on this image.

• In pose tracking, the input is usually a video (or real-time camera). The task is not only to detect the skeleton joint in this frame but also judge whether the skeleton joint belongs to the same identification as last frame.

2. Prior knowledge.

• Pose estimation: Nothing more than an RGB image itself.

• Pose tracking: Full video, and the initial skeleton joints of the first frame.

The above is just as my understanding. If you think I said something wrong, please kindly correct me.

[XiangpengZeng]

好的，非常高兴能够收到你的回复，我可否理解为，尽管pose estimation可以应用在视频中，但他始终只是在对每一帧进行单独的姿态估计；而pose tracking则不同，他除了需要在每一帧中进行pose的估计（每一帧仍然需要pose的估计），还需要判断当前帧中的pose和之前帧中的pose是否从属于同一个对象（同一个人），相当于在时序上有一个相似性比较以判断是否为同一个id？ 此外，可否推荐几篇有介绍pose tracking的论文以方便进一步理解。再次感谢你之前的解答，十分感激！

[jiang-du]

是的，这两个任务的关系大致是这样。"pose tracking"中，虽然说"每一帧仍然需要pose的估计"，但是因为有了上一个frame的pose点位置先验，因此预测难度会更低的。

关于pose tracking的论文，可以看看“PoseTrack: A Benchmark for Human Pose Estimation and Tracking”, arxiv 2017和"Simple Baselines for Human Pose Estimation and Tracking", ECCV 2018等文章