3D Human Pose Estimation section of the ECCV 2018 PoseTrack Challenge


Testing

  • • The data is organized as 24416 RGB images (JPG format) indexed from 00001.jpg to 24416.jpg

  • • You must predict 3x17 joints for the person present in each of the test images

  • • The steps are:

    • 1. Create a .JSON prediction file that has the following dictionary entries:

    • {"image_id": 1234, "keypoints": [0.0, 0.0, 0.0, -117.24859619140625, -19.70220947265625, 45.1982421875, -108.88580322265625, 403.3277282714844, 143.09912109375, -61.47943115234375, 799.3109741210938, 348.869140625, 117.24822998046875, 19.7021484375, -45.1982421875, 98.47625732421875, 435.5586242675781, 78.5751953125, 98.3531494140625, 838.9680786132812, 275.17724609375, 40.0106201171875, -242.6392822265625, -6.53857421875, 4.38812255859375, -472.4635009765625, -93.96826171875, -1.90411376953125, -542.703125, -154.216796875, 17.11083984375, -655.918212890625, -161.00830078125, 112.40789794921875, -432.13427734375,-141.9326171875, 221.29547119140625, -270.84747314453125, -313.99267578125, 15.122314453125, -342.2911376953125, -426.595703125, -111.588623046875, -459.7628173828125, -49.43017578125, -344.7430419921875, -352.4302978515625, -89.3369140625, -276.22674560546875, -382.2840576171875, -323.22802734375]}

      This is an example for an element, where "image_id" is the id (from 1 to 24416) of the image that is being tested on. The entry "keypoints" are your 3x17 (51) predictions (x0 y0 z0 x1 y1 z1 ... x17 y17 z17) for the pose.

      An example .JSON is EXAMPLE.JSON in the Scripts folder.

    • 2. Run the json2binary.py script provided in the Scripts folder as following:

    • MY_PATH/Scripts/json2binary.py -j PATH_TO_JSON/SUBMISSION.JSON -b YOUR_SUBMISSION_NAME

    • 3. Go to the "Results" tab on the website, select the CHALL_H80K testset, browse for your CHALL_H80K_YOUR_SUBMISSION_NAME.results file and click "Upload".

    • 4. The entry in the "Results" tab, under CHALL_H80K testset, will have the descriptions "CHALL_H80K_YOUR_SUBMISSION_NAME" which you can edit in the associated text box.

    • 5. When you are ready to make your results public, press "Publish!" and you should see your results in the "Ranking" tab in the CHALL_H80K testset.