[One-page summary] Templates for 3D Object Pose Estimation Revisited: Generalization to New objects and Robustness to Occlusions (CVPR 2022) by Nguyen et al.

● Summary: Template-based new object and occlusion robust 3D pose estimation by contrastive learning

 

● Approach highlight

• Local representation(Unseen object performance): masking global feature with binary template mask M to solve the problem of unseen object's cluttered backgrounds

• Occlusion mask(Occlusion robustness): use a similarity-based occlusion mask O instead of a pooling layer to preserve important information

●  Main results

 

● Discussion

• Acc15 can't catch fine grained pose errors.
• I guess it is still a poor performance compared to the SoTA pose estimation model.