campose maybe is not corret

[Traffic-X]

I use the google earth dataset for SFM and transform the cam pose to Colmap format, but I found some mistake, can you give me some advice, process is below
First, I run [dateset_generator.py] to process Google Earth, then the [CameraPoses.csv] can be generated
Second, I read the code in [dateset_generator.py] , cam pose in .csv means Camera2World transform, so I should change it to W2C following Colmap
Finally, I read the code in [inference.py] and [diff_gaussian_rasterization/init.py], I think W2C is also needed for Gaussian render too, then I found "_get_w2c_matrix(self, cam_position, cam_quaternion)" function, I use the same code for Colmap format, but it doesn't work well. For SFM, I use Triangulator directly, but the reconstruction result is wrong.

So I think the cam coord is maybe wrong and also a question for gaussian render, can you give me some advice,

[Traffic-X]

Use colmap recalib campose without initialization and run spare reconstruciton

[hzxie]

The camera pose must be correct. You can uncomment the following code to verify the camera poses.

GaussianCity/scripts/dataset_generator.py

Lines 1621 to 1653 in c2c31ab

	# # Debug: Render Instance Map with Gaussian Splatting
	# # Revert the cx value for the KITTI 360 dataset. Otherwise, the image cannot be aligned perfectly.
	# _cam_rig = copy.deepcopy(cam_rig)
	# if dataset == "KITTI_360":
	# _cam_rig["intrinsics"][2] = (
	# cam_rig["sensor_size"][0] - cam_rig["intrinsics"][2]
	# )
	# scales = scales[vp_idx]
	# import extensions.diff_gaussian_rasterization as dgr
	# gr = dgr.GaussianRasterizerWrapper(
	# np.array(_cam_rig["intrinsics"], dtype=np.float32).reshape((3, 3)),
	# _cam_rig["sensor_size"],
	# flip_lr=True,
	# flip_ud=dataset == "KITTI_360",
	# device=torch.device("cuda"),
	# )
	# with torch.no_grad():
	# gs_points = utils.helpers.get_gaussian_points(
	# torch.from_numpy(points[:, :3]).unsqueeze(0),
	# scales.unsqueeze(0) * 0.5,
	# {
	# "rgb": torch.from_numpy(
	# utils.helpers.get_ins_colors(points[:, 4])
	# ).unsqueeze(0)
	# / 255.0
	# },
	# ).cuda()
	# img = gr(gs_points.squeeze(0), cam_pos, cam_quat)
	# cv2.imwrite(
	# "output/test.jpg",
	# img.permute(1, 2, 0).cpu().numpy()[..., ::-1] * 255,
	# )