About how to apply this work on my own data?

[ZHyuwind]

First, thanks for your work. I want to apply your works on my own data, please tell me how to do?Thanks again!

[mks0601]

You can see this code. Once you convert your dataset to MSCOCO format, you can easily adapt that code to use your dataset.

[ZHyuwind]

Thank you very much. I will try. 

…

---Original--- From: "Gyeongsik Moon"<notifications@github.com&gt; Date: Tue, Dec 24, 2019 14:24 PM To: "mks0601/PoseFix_RELEASE"<PoseFix_RELEASE@noreply.github.com&gt;; Cc: "zy"<1453012267@qq.com&gt;;"Author"<author@noreply.github.com&gt;; Subject: Re: [mks0601/PoseFix_RELEASE] About how to apply this work on my own data? (#36) You can see this code. Once you convert your dataset to MSCOCO format, you can easily adapt that code to use your dataset. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[ZHyuwind]

Hello, I also meet a problem that my own dataset how to get annotations? or this file is not needed?Thanks.

[mks0601]

I cannot understand what you asked

[ZHyuwind]

I cannot understand what you asked

sorry，maybe I need to understand how to transform dataset format again. I feel so sorry about disturbing you. Happy Christmas Eve！

[ZHyuwind]

Hello, I want to ask a question that my own data also need annotations file like COCO's person_keypoints_val2017.json?

[mks0601]

Right. You should convert annotation of your own dataset to MSCOCO format.

[mks0601]

Maybe these codes would be helpful.
https://github.com/mks0601/TF-SimpleHumanPose/tree/master/tool

These codes convert mpii or posetrack datasets to coco format.

[ZHyuwind]

Thanks a lot, I will try!

[ZHyuwind]

Hello, I used coco2017val dataset to get results from TF-simple pose, then take it as input_pose.json, but also appear "FileNotFoundError: [Errno 2] No such file or directory: 'tmp_result_0.pkl" ,I don't know how to deal with it. Please give me some advices, thank you very much!

[mks0601]

The tmp_result_0.pkl is a just temporal file, which is automatically generated during testing. You should set proper dataset and testset in main/config.py and proper path in data/COCO/dataset.py.

[ZHyuwind]

The tmp_result_0.pkl is a just temporal file, which is automatically generated during testing. You should set proper dataset and testset in main/config.py and proper path in data/COCO/dataset.py.

Thank you, my problem has solved.

[lxy5513]

@mks0601 Hi, Could you please tell me how to generate /test_on_trainset/result.json in my own dataset.
Could I produce it by TF-SimpleHumanPose?, It is indispensable? and what's it role?

[mks0601]

Most of the images do not have all keypoint coordinate annotations because of truncations or severe occlusions. I used test_on_trainset to fill the not annotated keypoint coordinates. The format of the .json file exactly follows MSCOCO format. You can use any human pose estimation method to get it.

[lxy5513]

@mks0601 Thanks your kind respond, Do you mean that original coordinate annotations may not include all keypoints, So the function of test_on_trainset is to mitigate the problem ?

[mks0601]

Yes. For example, the annotation file of COCO dataset contains keypoints like the below figure.

Among x1, y1, v1, ..., v1 provides the below information.

Each keypoint has a 0-indexed location x,y and a visibility flag v defined as v=0: not labeled (in which case x=y=0), v=1: labeled but not visible, and v=2: labeled and visible.

If v==0, the coordinates are not annotated, which requires test_on_trainset.

[lxy5513]

So kind of you, thanks!