Small array issue in data process

[tshadley]

Looking at

part-affinity/src/data_process/coco_process_utils.py

Line 48 in 0f49c68

out_map[i] = np.maximum(out_map[i], DrawGaussian(out_map[i], keypoint[0:2], sigma=sigma))

I think there's a subtle bug here in that DrawGaussian will overwrite out_map[i] before np.maximum can evaluate the prior array. A quick fix is:

out_map[i] = np.maximum(out_map[i].copy(), DrawGaussian(out_map[i], keypoint[0:2], sigma=sigma))

This will affect any training data where two or more person pose points overlap within 43 pixels (given current gaussian settings). I don't know how it affects performance or loss though.