Research map / Detection and segmentation / 3D detection and driving
Semi-Open 3D Object Retrieval via Hierarchical Equilibrium on Hypergraph
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the semi-open 3D object retrieval setting with hierarchical labels, four derived datasets, and the HERT framework (hierarchical retrace embedding plus hypergraph-based structured equilibrium tuning) that outperforms prior retrieval and open-set baselines.
Problem
Open-set 3D object retrieval assumes single-level labels and no overlap between training and testing categories, which causes contradictory optimization when categories share coarse labels but differ in fine labels; the paper targets a semi-open setting where coarse label spaces overlap and fine lab
Approach
Combine hierarchical retrace embeddings (coarse-label-conditioned auto-encoder) with a superposed hypergraph whose hyperedges encode local coherent correlations (same coarse label) and global entangled correlations (kNN), followed by hypergraph convolution and a memory bank to generalize to unseen fine categories.
Key results
- On SO-NTU, HERT reaches mAP 0.5678 (about 17.7% over the second-best method), per the paper.
- On SO-MN40, HERT reaches mAP 0.6336 and Recall 0.3993 (about 9.6% and 7.9% over second-best HGM2R).
- Ablations show removing the retrace encoding drops SO-MN40 mAP from 0.6336 to 0.5791, and replacing hypergraph convolution with GCN or MLP lowers performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HERT instead of HGM2R | Semi-open 3D object retrieval · SO-ESB | mAP: 0.5049 -> 0.5756 | Helped |
| Use HERT instead of HGM2R | Semi-open 3D object retrieval · SO-NTU | mAP: 0.4821 -> 0.5678 | Helped |
| Use HERT instead of HGM2R | Semi-open 3D object retrieval · SO-MN40 | mAP: 0.5779 -> 0.6336 | Helped |
| Use HERT instead of HGM2R | Semi-open 3D object retrieval · SO-ABO | mAP: 0.6069 -> 0.6339 | Helped |
| Add retrace encoding (ablation) | HERT HRE · SO-ESB, SO-NTU, SO-MN40, SO-ABO | mAP: 0.5159 -> 0.5756; 0.4913 -> 0.5678; 0.5791 -> 0.6336; 0.6055 -> 0.6339 | Helped |
| Add HRE cross-entropy loss (ablation) | HERT HRE · SO-ESB, SO-NTU, SO-MN40, SO-ABO | mAP: 0.5133 -> 0.5756; 0.5161 -> 0.5678; 0.5967 -> 0.6336; 0.5885 -> 0.6339 | Helped |
| Add coherent hyperedges (ablation) | HERT SET · SO-ESB, SO-NTU, SO-MN40, SO-ABO | mAP: 0.5358 -> 0.5756; 0.5285 -> 0.5678; 0.5913 -> 0.6336; 0.6006 -> 0.6339 | Helped |
| Use hypergraph convolution instead of GCN (ablation) | HERT SET · SO-ESB, SO-NTU, SO-MN40, SO-ABO | mAP: 0.5405 -> 0.5756; 0.5144 -> 0.5678; 0.5602 -> 0.6336; 0.5686 -> 0.6339 | Helped |
| Use hypergraph convolution instead of MLP (ablation) | HERT SET · SO-ESB, SO-NTU, SO-MN40, SO-ABO | mAP: 0.5014 -> 0.5756; 0.4689 -> 0.5678; 0.5088 -> 0.6336; 0.4880 -> 0.6339 | Helped |
Limitations
- Only three levels of labels could be verified due to dataset limitations.
- Datasets are derived from existing 3D datasets with coarse labels assigned by basic shape and some objects removed.
- Code, models, and datasets are not yet released.
Experimental setup
Datasets: SO-ESB SO-NTU SO-MN40 SO-ABO
Benchmarks: SO-ESB SO-NTU SO-MN40 SO-ABO
Metrics: mAP Recall NDCG ANMRR precision-recall curve
Baselines: SDML CMCL MMSAE TranGCN C2AE HGM2R
Builds on: HGNN HGM2R TranGCN C2AE
Closest papers by meaning
- Assembly Fuzzy Representation on Hypergraph for Open-Set 3D Object Retrieval (NeurIPS 2024 · similarity 0.89)
- Towards 3D Objectness Learning in an Open World (NeurIPS 2025 · similarity 0.84)
- 3DCoMPaT200: Language Grounded Large-Scale 3D Vision Dataset for Compositional Recognition (NeurIPS 2024 · similarity 0.84)
- Open3DSG: Open-Vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-Set Relationships (CVPR 2024 · similarity 0.84)
- Uni3D: Exploring Unified 3D Representation at Scale (ICLR 2024 · similarity 0.83)
- Cross-Modal 3D Representation with Multi-View Images and Point Clouds (CVPR 2025 · similarity 0.83)
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.