about camera extrinsic

[duruing]

Hi! I'm reading your code recently and I'm confused why don't you reverse the y-axis and z-axis here when converting extrinsic from opencv coordinate system to open3d coordinate system

2d-gaussian-splatting/utils/mesh_utils.py

Lines 64 to 66 in 6e21151

	extrinsic=np.asarray((viewpoint_cam.world_view_transform.T).cpu().numpy())
	camera = o3d.camera.PinholeCameraParameters()
	camera.extrinsic = extrinsic

I notice that you implement this in estimate_bounding_sphere function

2d-gaussian-splatting/utils/mesh_utils.py

Lines 125 to 132 in 6e21151

	def estimate_bounding_sphere(self):
	"""
	Estimate the bounding sphere given camera pose
	"""
	from utils.render_utils import transform_poses_pca, focus_point_fn
	torch.cuda.empty_cache()
	c2ws = np.array([np.linalg.inv(np.asarray((cam.world_view_transform.T).cpu().numpy())) for cam in self.viewpoint_stack])
	poses = c2ws[:,:3,:] @ np.diag([1, -1, -1, 1])

I'm confused about the difference. Could you help explain? Thanks!

[hbb1]

The poses are transformed to opengl format because the function for computing focus point uses that format.