Thanks for share your code.
Correlation between person centers and thorax predictions in prepareTestResults.m
0.9893
0.9083
What's the use of two above?

Hi,

Thanks for sharing your code. Do you run the https://github.com/anewell/pose-hg-demo/ to get he anewell's prediction on the validation set? If so, I think the comparison is NOT totally comparable. Because the trick in file https://github.com/anewell/pose-hg-train/blob/master/src/util/pose.lua

-- Very simple post-processing step to improve performance at tight PCK thresholds for i = 1,p:size(1) do for j = 1,p:size(2) do local hm = tmpOutput[i][j] local pX,pY = p[i][j][1], p[i][j][2] scores[i][j] = hm[pY][pX] if pX > 1 and pX < opt.outputRes and pY > 1 and pY < opt.outputRes then local diff = torch.Tensor({hm[pY][pX+1]-hm[pY][pX-1], hm[pY+1][pX]-hm[pY-1][pX]}) p[i][j]:add(diff:sign():mul(.25)) end end end p:add(0.5)

is not used in the getPreds function of file https://github.com/anewell/pose-hg-demo/blob/master/util.lua
. Besides, the transform function in https://github.com/anewell/pose-hg-train/blob/master/src/util/img.lua is kind of different from transform function in https://github.com/anewell/pose-hg-demo/blob/master/img.lua.

The bearpaw's prediction are based on the code in https://github.com/anewell/pose-hg-train rather than https://github.com/anewell/pose-hg-demo, which should have more advantages.

I would like to ask about the submission of MPII test set. Besides the test set prediction results (pred_keypoints_mpii. Mat) that mpi require, should I prepare any other relevant reports?

what is the order of the joints in the prediction files?

I use the test.json(7000+ samples) provided by HRNet to generated pred.mat, then i use your code to convert it to pred_mpii_keypoints.mat, but in this process, there rise a error that the numbers of sample is not same 10000+. In other words, the test.json provided by HRNet does not include all test samples. So i use the official test.h5 to prepare the pred_mpii_keypoints.mat. But i find there is a difference between test.json(70000)+ and tets.h5(10000) + in the items of center and scale. does this have a significant influence about testing? Or can i directly use test.h5 to submit?

Hi @anibali and thanks for your work!

Can you share the script for converting the mpii_human_pose_v1_u12_1.mat to the .h5 files? Do you know if there is code for converting the mpii_human_pose_v1_u12_1.mat to a .csv file?
Im asking because I want to train my model explicity on one category of movement (act_id) within MPII dataset. So im thinking of converting the dataset to .csv file i order to delete the not desired data manually. Then i want to bring it in this .csv format or this .h5 format. What do you think about my idea? Any hints for solving that?

Hi @anibali ,

I saw their codes you referenced for plotting, sth confusing is that for validation and testing they used the same index from MPII and as it is mentioned in stacked hourglass paper they provide results on validation because test annotation is not available , and you have provided annotation folder which separate train, val, test index, I confused a little, would you please clarify me?

Hi,
You are stating that this repo is only for single person evaluation. What confuses me is the fact that the test.h5 database contains 11,731 persons, while the official mpii_human_pose_v1_u12_1.mat file has a total of 11,823 persons annotated as test. Of these, only 7,247 are annotated as single persons. So, I am wondering which format the submission would actually need. I find the documentation quite sparse, but if I understood this (http://human-pose.mpi-inf.mpg.de/#evaluation) correctly, only the 7,247 persons should be predicted in the single-person case.
Could you explain why there are 11,731 persons annotated in test.h5?

hi, anibail, how to generate his final result of 90.09 in test set with https://github.com/bearpaw/pytorch-pose
?And how to generate the test set? thanks

Hi,

Thanks for sharing your code.　I want to evaluate our dataset on single keypoint. but met some issues, I find the number of gt must be match with dt's? Thanks very much.
｀｀｀
uv_error = pos_pred_src - pos_gt_src
ValueError: operands could not be broadcast together with shapes (14,2,1743) (1917,14,2)
｀｀｀

你好 pckh什么意思呢？？？？比如DNN得到预测的一堆xy坐标 和ground truth的xy直接 怎样计算呢？？

help

Scripts
Loads predictions from Matlab or HDF5 files and compares them with ground truth labels to calculate accuracy metrics (eg PCKh). You will want to edit getExpParamsNew.m to add new sets of predictions, and evalMPII.m to specify which predictions to include and which subset (train/val) to use.

I have edited getExpParamsNew.m to add the new prediction such as

case 6
    p.name = 'XXX';
    p.predFilename = 'preds/reference/XXX.mat';

And I edited the evalMPII.m

PRED_IDS = [1, 2, 5, 6];

But it shows the error
Error using evalMPII (line 82)
Assertion failed.

I want to evaulate my test_pred.mat, and send an email to [email protected], debut no one answered my email.

When we do Multi person Estimation and generate the prediction mat

% Set predicted x_pred, y_pred coordinates and prediction score for each body joint
pred(imgidx).annorect(ridx).annopoints.point(pidx).x = x_pred;
pred(imgidx).annorect(ridx).annopoints.point(pidx).y = y_pred;
pred(imgidx).annorect(ridx).annopoints.point(pidx).score = score;

Should ridx be in rectidxs_multi_test (extracted by getMultiPersonGroups(groups,RELEASE,false))? Or it's just a iterator over the people in an image according to my preditions result?

hi, i want to know what is the matrix shape of pred_keypoints_mpii.mat when you submit to mpii.

prepareTestResults (line 42)
assert fail

I have generate a (11731,16,2) sized test_preds.h5, and i want to transform it to official mat file. Have you did it before that submit a predict result to official mail and get a feedback result. @anibali hope for your help~