norm = np.ones((pred.shape[0], 2)) * np.array([h, w]) / 10

[wonss737]

In the accuracy function in evaluate.py, predicted keypoint coordinate is normalized by [h,w] / 10. For example, if heatmap size is (64,48), normalization factor is (64,48).

But, the 'pred', which is the output of 'get_max_preds' function, has [w,h] scale. (e.g, [34, 58], [40, 50]). So I think the normalization factor should be [w, h].

Thank you :)