pose-evaluation
pose-evaluation copied to clipboard
About AKD
Hi, @AliaksandrSiarohin . In your paper, you discribed that AKD is obtained by computing the average distance between the detected keypoints of the ground truth and of the generated video. I have read your code. I am curious about the average distance. Is it the value of the absolute coordinate in x, y. or the value of ratio of coordinate divides the image size.
since in your paper, that number is only 1.119 of dataset nemo, 1.294 of dataset Voxceleb. which means only about 1 pixels average shift of all the face keypoints?
In your code:
scores.append(np.sum(np.abs(df1['value'].iloc[i] - df2['value'].iloc[i]).astype(float) ** 2))
maybe it should be this?
scores.append(np.sum(np.abs(df1['value'].iloc[i] - df2['value'].iloc[i]).astype(float) ** 2) / (image.size[0]*image.size[1]))