import roma import torch import torch.nn.functional as F from torch.amp import autocast def quaternion_multiply(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor: """Composition of rotations given as wxyz quaternions, returned with a non-negative real part.""" aw, ax, ay, az = torch.unbind(a, -1) bw, bx, by, bz = torch.unbind(b, -1) ab = torch.stack(( aw * bw - ax * bx - ay * by - az * bz, aw * bx + ax * bw + ay * bz - az * by, aw * by - ax * bz + ay * bw + az * bx, aw * bz + ax * by - ay * bx + az * bw, ), -1) return torch.where(ab[..., 0:1] < 0, -ab, ab) @torch.no_grad() @autocast("cuda", enabled=False) def transform_gaussians_to_cam_coord(c2w: torch.Tensor, gaussians: torch.Tensor) -> torch.Tensor: """Gaussians [B, N, 14] (pos, scale, quat wxyz, ...) -> the same Gaussians in the camera frame of the rigid camera-to-world matrices c2w [B, 4, 4].""" w2c = roma.Rigid.from_homogeneous(c2w).inverse() pos = w2c[:, None].apply(gaussians[..., :3]) w2c_quat = roma.rotmat_to_unitquat(w2c.linear)[..., [3, 0, 1, 2]] # roma is xyzw rot = F.normalize(gaussians[..., 6:10], dim=-1) # padded all-zero rows stay zero instead of NaN rot = quaternion_multiply(w2c_quat[:, None], rot) return torch.cat([pos, gaussians[..., 3:6], rot, gaussians[..., 10:]], dim=-1)