# either extract median extrinsic or just most confident one (using weighted mean here) 
            # just trying max idx
            #est_extrinsics = est_links[ torch.tensor([x.sum() for x in segs]).max(dim=0)[1].item() ]
            #weights = np.array([x.sum() for x in segs])
            #weights = weights/weights.max()
            #try: est_extrinsics=geometry.average_se3_split(est_links,weights)
            #except: print("bad se3 est");  est_extrinsics=np.eye(4)
            #print((np.linalg.inv(model_input["cam2world_cv"].cpu().numpy())-est_extrinsics).round(2)) # should be 0 for GT
            # try procrustes on all points and then with ransac procrustes