Table 1: The results of a single 3D object reconstruction presented by mean intersection over union Mean-IoU and 3D-Angular-Error calculated for the CUB test set when predicting camera pose by four different approaches, which are unit quaternions, Gram-Schmidt, special orthogonalization, and the keypoint prediction. Note that the category-specific mesh reconstruction network CMR had a 3D-Angular error equal to 87.52° when no viewpoint and keypoint supervision are used.
Method Mean-IoU ↑ 3D-Angular-Error ↓
Quaternion Goel et. al (2020) 0.62 45.5°
Gram-Schmidt Zhou et. al (2020) 0.55 60.4°
SVD Levinson et. al (2020) 0.58 50.8°
Keypoint (ours) 0.70 40.5°