Open Darius-H opened 1 year ago
I find that when using pytorch3d, convert_screen_to_ndc function does not work properly, it generates a wrong principal point. The convert_screen_to_ndc is here: https://github.com/open-mmlab/mmhuman3d/blob/9431addec32f7fbeffa1786927a854c0ab79d9ea/mmhuman3d/core/conventions/cameras/convert_convention.py#L481C1-L481C1 I find that in pytorch3d’s code , there is a principal point fix matrix, which is not included in mmhuman's code: https://github.com/facebookresearch/pytorch3d/blob/57f6e79280e78b6e8308f750e64d32984ddeaba4/pytorch3d/renderer/cameras.py#L1187C1-L1224C1
I find that when using pytorch3d, convert_screen_to_ndc function does not work properly, it generates a wrong principal point. The convert_screen_to_ndc is here: https://github.com/open-mmlab/mmhuman3d/blob/9431addec32f7fbeffa1786927a854c0ab79d9ea/mmhuman3d/core/conventions/cameras/convert_convention.py#L481C1-L481C1 I find that in pytorch3d’s code , there is a principal point fix matrix, which is not included in mmhuman's code: https://github.com/facebookresearch/pytorch3d/blob/57f6e79280e78b6e8308f750e64d32984ddeaba4/pytorch3d/renderer/cameras.py#L1187C1-L1224C1