This README provides instructions on how to prepare your predictions using MATLAB for evaluation on MPII HumanPose Dataset. Predictions should emailed to [eldar at mpi-inf.mpg.de].

At most four submissions for the same approach are allowed. Any two submissions must be 72 hours apart.

Preliminaries

Download evaluation toolkit.

Multi-Person Pose Estimation

Evaluation protocol

• Evaluation is performed on groups of multiple people. One image may contain multiple groups.
• Each group is localized by computing x1, y1, x2, y2 group boundaries from locations of all people in the group and cropping around those boundaries.

    pos = zeros(length(rect),2);   for ridx = 1:length(rect)       pos(ridx,:) = [rect(ridx).objpos.x rect(ridx).objpos.y];   end   x1 = min(pos(:,1)); y1 = min(pos(:,2)); x2 = max(pos(:,1)); y2 = max(pos(:,2));

• Scale of each group is computed as an average scale of all people in the group.

    scale = zeros(length(rect),2);   for ridx = 1:length(rect)       scale(ridx) = rect(ridx).scale;        end   scaleGroup = mean(scale);

• Using ground truth number of people is not allowed.
• Using approximate location of each person while estimating person's pose is not allowed.
• Using individual scale of each person while estimating person's pose is not allowed.

Preparing predictions

1. Extract testing annotation list structure from the entire annotation list

    annolist_test = RELEASE.annolist(RELEASE.img_train == 0);

2. Extract groups of people using getMultiPersonGroups.m function from evaluation toolkit

    load('groups_v12.mat','groups');  [imgidxs_multi_test,rectidxs_multi_test] = getMultiPersonGroups(groups,RELEASE,false);

   where imgidxs_multi_test are image IDs containing groups and rectidxs_multi_test are rectangle IDs of people in each group.

3. Split testing images into groups

    pred = annolist_test(imgidxs_multi_test);

4. Set predicted x_pred, y_pred coordinates in the original image, 0-based joint id (see 'annotation description') 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).id = id;  pred(imgidx).annorect(ridx).annopoints.point(pidx).score = score;

5. Save predictions into pred_keypoints_mpii_multi.mat

    save('pred_keypoints_mpii_multi.mat','pred');

Evaluation Script

Evaluation is performed by using evaluateAP.m function

Single Person Pose Estimation

Evaluation protocol

• Evaluation is performed on sufficiently separated people.
• Using approximate location and scale of each person is allowed

    pos = [rect(ridx).objpos.x rect(ridx).objpos.y];   scale = rect(ridx).scale;

Preparing predictions

1. Extract testing annotation list structure from the entire annotation list:

    annolist_test = annolist(RELEASE.img_train == 0);

2. Extract image IDs and rectangle IDs of single persons

    rectidxs = RELEASE.single_person(RELEASE.img_train == 0);

3. Set predicted x_pred, y_pred coordinates in the original image for each body joint of single persons and 0-based joint id (see 'annotation description')

    pred = annolist_test;  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).id = id;

4. Save predictions into pred_keypoints_mpii.mat

    save('pred_keypoints_mpii.mat','pred');

Evaluation Script

Evaluation is performed by using evaluatePCKh.m function