AI Paper Atlas

Research map / 3D vision / Human pose estimation

Coupled Iterative Refinement for 6D Multi-Object Pose Estimation

Lahav Lipson, Zachary Teed, Ankit Goyal, Jia Deng

CVPR 2022 · 3D vision › Human pose estimation · approach: Instance-level 6D object pose · 67 citations (OpenAlex)

Check your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Use coupled iterative pose-and-correspondence refinement instead of one-shot flow followed by PnP (ablation)Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSSD Recall: 0.643 -> 0.955 (YCB-V); 0.275 -> 0.582 (T-LESS); 0.054 -> 0.788 (LM-O)Improved
Use bidirectional rather than single-direction correspondences (ablation)BD-PnP · YCB-V, T-LESS, LM-OMSPD Recall: 0.905 -> 0.924 (YCB-V); 0.677 -> 0.685 (T-LESS); 0.605 -> 0.828 (LM-O)Improved
Add depth augmentation and predict depth revisions (ablation)BD-PnP · YCB-V, T-LESS, LM-OMSPD Recall: 0.909 -> 0.924 (YCB-V); 0.678 -> 0.685 (T-LESS); 0.819 -> 0.828 (LM-O)Improved
Predict per-pixel confidence weights rather than use uniform confidence (ablation)Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSSD Recall: 0.832 -> 0.955 (YCB-V); 0.424 -> 0.582 (T-LESS); 0.760 -> 0.788 (LM-O)Improved
Use multiple perturbed renders rather than a single render (ablation)Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSSD Recall: 0.941 -> 0.955 (YCB-V); 0.545 -> 0.582 (T-LESS); 0.641 -> 0.788 (LM-O)Improved
Train with both pose and flow losses rather than either alone (ablation)Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSPD Recall: 0.740 -> 0.924 and 0.866 -> 0.924 (YCB-V); 0.558 -> 0.685 and 0.261 -> 0.685 (T-LESS); 0.804 -> 0.828 and 0.615 -> 0.828Improved
Use four outer loops instead of one (ablation)Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSPD Recall: 0.924 -> 0.933 (YCB-V); 0.685 -> 0.694 (T-LESS); 0.828 -> 0.831 (LM-O)Improved
Refine the initial pose (ablation)Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSPD Recall: 0.194 -> 0.924 (YCB-V); 0.263 -> 0.685 (T-LESS); 0.475 -> 0.828 (LM-O)Improved
Treat rendered depth as a variable rather than a constant for RGB input (ablation)RGB Coupled Iterative Refinement · YCB-V, T-LESSMSSD Recall: 0.678 -> 0.751 (YCB-V); 0.343 -> 0.474 (T-LESS)Improved
Discard the optimized depth update and regenerate depth from the updated pose (ablation)RGB Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSSD Recall: 0.674 -> 0.751 (YCB-V); 0.326 -> 0.474 (T-LESS); 0.497 -> 0.609 (LM-O)Improved
Use bidirectional context features rather than unidirectional context features (ablation)Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSPD Recall: 0.848 -> 0.833 (YCB-V RGB); 0.647 -> 0.649 (T-LESS RGB); 0.807 -> 0.793 (LM-O RGB); 0.910 -> 0.924 (YCB-V RGB-D)Mixed
Add depth revisions in the RGB-only setting (ablation)RGB Coupled Iterative Refinement · YCB-V, T-LESS, LM-OMSPD Recall: 0.833 -> 0.842 (YCB-V); 0.649 -> 0.647 (T-LESS); 0.793 -> 0.803 (LM-O)Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy