fix(torch-transform): inverse_transform computes wrong (+1 more)

[andrewwhitecdw]

Small fixes in lib/utils/torch_transform.py:

fix: inverse_transform computes wrong inverse translation

Fix: Replace:

    transform_inv[..., :3, 3] = -torch.matmul(transform_mat[..., :3, 3].unsqueeze(-2), transform_mat[..., :3, :3]).squeeze(-2)

with:

    transform_inv[..., :3, 3] = -torch.matmul(transform_mat[..., :3, :3].transpose(-2, -1), transform_mat[..., :3, 3].unsqueeze(-1)).squeeze(-1)

fix: quaternion helper uses hard-coded float64 dtype

Fix: Replace:

    out[sind] = torch.tensor([1., 0., 0., 0.], device=v1.device)
    # handle v1 & v2 with opposite direction
    nind = dot < -1 + eps
    if torch.any(nind):
        vx = torch.tensor([1., 0., 0.], device=v1.device)
        vxdot = (v1 * vx).sum(-1).abs()
        nxind = nind & (vxdot < 1 - eps)
        if torch.any(nxind):
            out[nxind] = angle_axis_to_quaternion(normalize(torch.cross(vx.expand_as(v1[nxind]), v1[nxind], dim=-1)) * np.pi)
        # handle v1 & v2 with opposite direction and they are parallel to x axis
        pind = nind & (vxdot >= 1 - eps)
        if torch.any(pind):
            vy = torch.tensor([0., 1., 0.], device=v1.device)
            out[pind] = angle_axis_to_quaternion(normalize(torch.cross(vy.expand_as(v1[pind]), v1[pind], dim=-1)) * np.pi)

with:

    out[sind] = torch.tensor([1., 0., 0., 0.], device=v1.device, dtype=v1.dtype)
    # handle v1 & v2 with opposite direction
    nind = dot < -1 + eps
    if torch.any(nind):
        vx = torch.tensor([1., 0., 0.], device=v1.device, dtype=v1.dtype)
        vxdot = (v1 * vx).sum(-1).abs()
        nxind = nind & (vxdot < 1 - eps)
        if torch.any(nxind):
            out[nxind] = angle_axis_to_quaternion(normalize(torch.cross(vx.expand_as(v1[nxind]), v1[nxind], dim=-1)) * np.pi)
        # handle v1 & v2 with opposite direction and they are parallel to x axis
        pind = nind & (vxdot >= 1 - eps)
        if torch.any(pind):
            vy = torch.tensor([0., 1., 0.], device=v1.device, dtype=v1.dtype)
            out[pind] = angle_axis_to_quaternion(normalize(torch.cross(vy.expand_as(v1[pind]), v1[pind], dim=-1)) * np.pi)

Files changed

• lib/utils/torch_transform.py