What is the use for "Debug Camera offset"?

Thanks for your great work!

I find that in the dataparser file:

    #####################
    # Debug Camera offset

There is a piece of code called "Debug Camera offset", which changes the normal pose reading and rotates it by a certain angle.
Will this affect the results? What is its usage?
And should I comment it out?

def get_camera_poses_tracking(poses_velo_w_tracking, tracking_calibration, selected_frames, scene_no=None):
    exp = False
    camera_poses = []

    opengl2kitti = np.array([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]])

    start_frame = selected_frames[0]
    end_frame = selected_frames[1]

    #####################
    # Debug Camera offset
    if scene_no == 2:
        yaw = np.deg2rad(0.7)  ## Affects camera rig roll: High --> counterclockwise
        pitch = np.deg2rad(-0.5)  ## Affects camera rig yaw: High --> Turn Right
        # pitch = np.deg2rad(-0.97)
        roll = np.deg2rad(0.9)  ## Affects camera rig pitch: High -->  up
        # roll = np.deg2rad(1.2)
    elif scene_no == 1:
        if exp:
            yaw = np.deg2rad(0.3)  ## Affects camera rig roll: High --> counterclockwise
            pitch = np.deg2rad(-0.6)  ## Affects camera rig yaw: High --> Turn Right
            # pitch = np.deg2rad(-0.97)
            roll = np.deg2rad(0.75)  ## Affects camera rig pitch: High -->  up
            # roll = np.deg2rad(1.2)
        else:
            yaw = np.deg2rad(0.5)  ## Affects camera rig roll: High --> counterclockwise
            pitch = np.deg2rad(-0.5)  ## Affects camera rig yaw: High --> Turn Right
            roll = np.deg2rad(0.75)  ## Affects camera rig pitch: High -->  up
    else:
        yaw = np.deg2rad(0.05)
        pitch = np.deg2rad(-0.75)
        # pitch = np.deg2rad(-0.97)
        roll = np.deg2rad(1.05)
        # roll = np.deg2rad(1.2)

    cam_debug = np.eye(4)
    cam_debug[:3, :3] = get_rotation(roll, pitch, yaw)

    Tr_cam2camrect = tracking_calibration["Tr_cam2camrect"]
    Tr_cam2camrect = np.matmul(Tr_cam2camrect, cam_debug)
    Tr_camrect2cam = invert_transformation(Tr_cam2camrect[:3, :3], Tr_cam2camrect[:3, 3])
    Tr_velo2cam = tracking_calibration["Tr_velo2cam"]
    Tr_cam2velo = invert_transformation(Tr_velo2cam[:3, :3], Tr_velo2cam[:3, 3])

    camera_poses_imu = []
    for cam in camera_ls:
        Tr_camrect2cam_i = tracking_calibration["Tr_camrect2cam0" + str(cam)]
        Tr_cam_i2camrect = invert_transformation(Tr_camrect2cam_i[:3, :3], Tr_camrect2cam_i[:3, 3])
        # transform camera axis from kitti to opengl for nerf:
        cam_i_camrect = np.matmul(Tr_cam_i2camrect, opengl2kitti)
        cam_i_cam0 = np.matmul(Tr_camrect2cam, cam_i_camrect)
        cam_i_velo = np.matmul(Tr_cam2velo, cam_i_cam0)

        cam_i_w = np.matmul(poses_velo_w_tracking, cam_i_velo)
        camera_poses_imu.append(cam_i_w)

    for i, cam in enumerate(camera_ls):
        for frame_no in range(start_frame, end_frame + 1):
            camera_poses.append(camera_poses_imu[i][frame_no])

    return np.array(camera_poses)

I take it a try. After I delete this part, the reconstruction quality of dynamic car will apparently decrease. (I try it in scene06)
Is there any relation?
(I just want to reconstruction the scene using the original poses)
Thanks a lot for your answer!
@xBeho1der @wuzirui

Here is the result after deleting this part:(Training for 1w5 Epoch)
Origin:

After deleting:

this is from the original nsg code. My guess is that the authors manually calibrated the noisy camera poses for better performance.

this is from the original nsg code. My guess is that the authors manually calibrated the noisy camera poses for better performance.

Thanks a lot

Also same file , (the highlighted text) how you calculating this

mars/mars/data/mars_kitti_dataparser.py

Line 1010 in 69b9bf9

t_crect2c = np.matmul(K_inv, R_t)