[Bug] T matrix incorrectly recomputed from P matrix , causing projection mismatch in full size img

[Kyekyekye]

Hi team,

I’ve been investigating the geometric projection pipeline in bevlab/datasets.py, and I noticed a possible issue regarding the computation of the camera translation vector T.

Specifically, when testing the projection of Velodyne points to the full-size image using the extracted R_data, T_data, and K_data, I found that the result does not match the expected projection defined by KITTI:

x = P @ Tr @ X

To isolate the issue:

• I set the augmentation noise to zero:

  pc, trans_dict = revtrans_translation(pc, trans_dict)
  pc, trans_dict = revtrans_rotation(pc, trans_dict)

• I also disabled any resizing or modification of the camera intrinsics:

  K[0, 0] *= img_w_ratio
  K[0, 2] *= img_w_ratio
  K[1, 1] *= img_h_ratio
  K[1, 2] *= img_h_ratio
  

Under these conditions, I would expect that:

K_data @ (R_data @ pc.T + T_data.reshape(3, 1))

should yield projections that align with x = P @ Tr @ X
However, I observe a consistent shift. Upon inspection, I suspect the issue lies in the following recomputation of T:

tz = T[2]
tx = (P[0, 3] - cx * tz) / fx
ty = (P[1, 3] - cy * tz) / fy
T = np.array([tx, ty, tz])

Would appreciate your confirmation on whether this tx/ty adjustment was intended, or if it may be removed for clarity and correctness.

Thanks!