Not sure if the repository is actively supported, but the keypoints that are returned by GAST-net are normalized to [-1, 1] by default. If we want to plot them on a video (using only the x, y axes) we need to un-normalize them. But if I simply reverse the normalization the 3D predictions are scaled incorrectly, and furthermore the 2D skeleton retreived is not correct. Is there any readymade method to reconvert the estimated 3D keypoints back to 2D projection on top of a 2D video? Or any correct way to un-normalize the keypoints for comparison with the original input keypoints so they can be quantitatively compared?
Not sure if the repository is actively supported, but the keypoints that are returned by GAST-net are normalized to [-1, 1] by default. If we want to plot them on a video (using only the x, y axes) we need to un-normalize them. But if I simply reverse the normalization the 3D predictions are scaled incorrectly, and furthermore the 2D skeleton retreived is not correct. Is there any readymade method to reconvert the estimated 3D keypoints back to 2D projection on top of a 2D video? Or any correct way to un-normalize the keypoints for comparison with the original input keypoints so they can be quantitatively compared?