Repositioning Code about coco-analyze HOT 2 CLOSED

That happens in the _find_kpt_errors function. In particular:

• Lines 326-346: The different type of localization error is identified for each individual keypoint in the specific detection being analyzed. Issue #2 contains additional details that might help you further understand.

• Lines 348-376: Based on the type of error identified for each keypoint an update is performed to its x and y coordinates so that the KS of that keypoint is increased.

A few key observations:

1. For a specific keypoint (i.e. left-eye, right-knee, etc.), the KS between a predicted keypoint location and the ground-truth location is the same over all possible coordinates with a fixed radial distance from the ground-truth location. I.e. if a prediction x is 5 pixels away from the ground-truth location o, it doesn't matter if it's to its left (x-o) or right (o-x), the KS will be the same. For that reason, when correcting a keypoint, I only adjust its x coordinate and leave the y coordinate the same as in the original prediction. This is arbitrary, and has no bearing whatsoever on the final result.

2. A keypoint can be in one and only one category between miss, swap, inversion, jitter, and good. This is fundamental for the working of the analysis because the corrected location must be unique. The variables good_kpts, jitt_kpts, miss_kpts, swap_kpts, inv_kpts are masks with dimension 17 (total number of keypoints) and never have overlap.

3. If a keypoint prediction is good its coordinates are not corrected. In fact, the mask good_kpts multiplies the original x coordinates of the detection and those values are assigned to the new corrected keypoint coordinates.

4. If a keypoint prediction is miss its coordinates are corrected by putting the new coordinates at a distance from the ground-truth location that has KS=0.5.

5. If a keypoint prediction is jitter its coordinates are corrected by putting the new coordinates at a distance from the ground-truth location that has KS=0.85.

6. If a keypoint prediction is inversion or swap its coordinates have a KS>=0.5 but with the wrong ground-truth (belonging to the same person in the case of inversion, or a different one in the case of swap). These errors are corrected by putting the new coordinates at a distance from the correct ground-truth location that has the same KS value that the previous prediction had with the wrong ground-truth.

Feel free to re-open the issue if you have further questions.

from coco-analyze.

Thank you very much.

from coco-analyze.