Camera intrinsics overwritten

[Byte247]

Hi, I am trying to understand the Fast BEV projection code. I noticed that in the NuScenesMultiView_Map_Dataset2 , or more precisely in the NuScenesMultiViewDataset the camera intrinsics are overritten with the identity matrix. Why is that done? In the following pipelines this is not set back to the original values.

new_info = dict(
            sample_idx=data_info['sample_idx'],
            img_prefix=[None] * n_cameras,
            img_info=[dict(filename=x) for x in data_info['img_filename']],
            lidar2img=dict(
                extrinsic=[tofloat(x) for x in data_info['lidar2img']],
                intrinsic=np.eye(4, dtype=np.float32),
                lidar2img_aug=data_info['lidar2img_aug'],
                lidar2img_extra=data_info['lidar2img_extra']
            )
        )

[arnoldfychen]

My understanding:

Because data_info['lidar2img']  is  lidar2img_rt, which is actually intrinsic@lidar2cam_rt.T in NuScenesDataset's get_data_info():
            viewpad = np.eye(4)
            viewpad[:intrinsic.shape[0], :intrinsic.shape[1]] = intrinsic
            lidar2img_rt = (viewpad @ lidar2cam_rt.T)

I guess the code owner wanted to re-establish the intrinsic item in lidar2img (used later in FastBEV's _compute_projection() for scaling) in NuScenesMultiViewDataset's get_data_info(), so, assigned it with an identity matrix, as you know, the value of a matrix doesn't change when the matrix multiply with an identity matrix, i.e., I@A = A