Open jk456445 opened 7 months ago
Author needs to confirm but from my understanding, the 0 and 1 are between [-1, 1] and the 2 is the keypoint confidence score.
I would also be glad to know the answer to the post's question. These are almost certainly coordinates in three-dimensional space. At first I thought that the skeleton was inscribed in a cube with a side of length 2 (from -1 to 1), but in one of the predictions I saw that the minimum value is less than -1 (-1.06). And it puzzled me :)
I believe you can find the code for the scaling in the following file: motionbert\lib\utils\utils_data.py functions: crop_scale() and crop_scale_3d()
It looks like the keypoints are scaled to [-1,1] based on the bounding box (min and max coordinates in all axes).
I have the 3d keypoints value of a picture, but I am curious about the unit of these values, whether it's in pixels or something else.![image](https://github.com/Walter0807/MotionBERT/assets/81965933/bddc23df-a55f-4431-a944-3cce6d1ada1b)