How do you process hm36 data?

[Peter654q]

I want to use the training of integral regression, but raise the OS error.

OSError: Fail to read /data/hm36/images/s_07_act_14_subact_02_ca_04/s_07_act_14_subact_02_ca_04_000320.jpg

May I ask that how do you process the hm36 data?
Can you provide the code?
Thanks!

[youngstu]

I have the same problem.
If providing the train data link, it would be better.

[kyang-06]

I wrote a piece of code and analyze what the name of input file means

's' means subject in hm3.6m and in the training s1,s5,s6,s7,s8 were used,
'act' means action, which contains 15 category in hm3.6m but here it was numbered from 02 to 16. still 15 cat totally. I guess that the cat 'ALL' occupies no.1
'sub' is subaction and each action has 2 subactions both in hm3.6m and in this repo
'ca' means camera, 4 in total
every 's_XX_act_XX_subact_XX_ca_XX_XXXXXX.jpg' lasts for 200 frames, but don't know the exact counting number beginning or finishing

[kyang-06]

hey guys, I found the pre-processed data provided by this project https://github.com/xingyizhou/pose-hg-3d has the same name format.
may be the answer

[Ferrnya]

I have the same problem, but I didn't find how to solve it.
I changed the human36 dataset path but the image path always to be /data/hm36/images/s_07_act_14_subact_02_ca_04/s_07_act_14_subact_02_ca_04_000320.jpg....
Do you find some ways to solve it?

[lck1201]

@Ferrnya hi, the image path is loaded from cache file, you should modify it based on your own condition.

[lck1201]

@Peter654q @youngstu Hi, we use internal processed hm36 dataset, so I am afraid that I cannot share with you. But, you can organize yours according to our naming rules.
s = subject
act = action
sub = sub-action
ca = camera
all index start from 1

[lck1201]

Hi all, we've released our processed HM3.6M annotation, the images are organized in the same structure. Plz check.

[Jeff-sjtu]

@lck1201 Hi! I found the camera_id in human3.6m dataset is annotated as [54138969, 55011271, 58860488, 60457274]. Did they stand for HM_camera_idx = [1, 2, 3, 4] respectively?

[Ferrnya]

@lck1201 Hi! I processed images by cropping from bounding box, so I have the image following.

s_11_act_02_subact_02_ca_03_000002.jpg
I used such images to train and test. I found that I have bad results.
Maybe I don't need to crop people from image?

[lck1201]

@Ferrnya Hi, the crop seems correct, we need this step. BTW, I think you should check annotation, data pre-processing, training process... and see whether learning curve is abnormal..