about __opencv_camera2pytorch3d in dataset/base.py

[lcc815]

Hi, dear author,

As #15 already asked, why does the rotation matrix need to be tranposed ? Even though you have given the link, I still don't get it. what does the comment of We also transpose R (opencv multiplies points from the opposite=left side). mean?

thanks.

[DRosemei]

Hi, as it says "opencv multiplies points from the opposite=left side", so R denotes "world2camera" transform and R need to be transposed to "camera2world". You can see here in our repo.

[lcc815]

谢谢回复。我用中文请教可以吗？感觉中文表达得更清楚一点。 我问了一下chatgpt，它是这样说的 请问是它说这样吗？ 另外我没明白这个地方为什么会涉及到opencv——pytorch3d中的PerspectiveCameras需要传入的R应该是world2camera吧，所以这个地方是不是只需要把数据集(nuScenes or kitti)中定义的camera坐标系和pytorch3d中定义的camera坐标系统一就可以了呢？为什么会和opencv有关系呢？ 谢谢~

[DRosemei]

1. 原始链接中，是从opencv坐标系转，opencv是左乘，所以P_camera = P_world @ W2C，它的输入时W2C变换，所以在传入pytorch_3D前需要先转置，变成C2W，也就是R_pytorch3d = R.clone().permute(0, 2, 1)。同时opencv坐标系x,y是和screen坐标系相反，所以需要反转一下，也就是R_pytorch3d[:, :, :2] = -1， T_pytorch3d[:, :2] = -1。
2. 本文直接传的W2C进来，所以直接做变换就好了
3. 所以PerspectiveCameras传入的是C2W+xy轴转换，本文是套了官方的函数做了转换，你也可以提前转换好

[lcc815]

懂了，感谢！！