Describe the bug

The function depth_to_3d_v2 accepts a camera matrix of the shape (*, 3, 3). However, depth_to_3d_v2 internally uses the unproject_meshgrid fuction that only supports a camera matrix of the shape (3, 3). When feeding a camera matrix of the shape (B, 3, 3) into depth_to_3d_v2, a runtime error occurs.

Reproduction steps

import torch
from kornia.geometry import depth_to_3d_v2

B, H, W = 4, 512, 512

disparity = torch.rand(B, H, W)
intrinsics = torch.tensor([[2262.52, 0.0, 1096.98], [0.0, 2265.3017905988554, 513.137], [0.0, 0.0, 1.0]])[None].repeat(B, 1, 1)

points_3d = depth_to_3d_v2(disparity, intrinsics)

Expected behavior

To fix this bug, either depth_to_3d_v2 should be restricted to a shape of (3, 3) or unproject_meshgrid should be extended to (*, 3, 3).

Environment

The latest Kornia version was used. The bug is independent of the platform.

Additional context

No response