Research map / Detection and segmentation / 3D detection and driving
Assembly Fuzzy Representation on Hypergraph for Open-Set 3D Object Retrieval
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.
Proposes HAFR, a hypergraph-based part-assembly fuzzy representation framework (HIConv, IAE, SFR with leverage hypergraph and memory bank) for open-set 3D object retrieval, and constructs three part-annotated OpenPart datasets.
Problem
Open-set 3D object retrieval for unseen categories lacks object-level labels; part-level shape commonalities across categories are underexploited, and models suffer distribution skew on unseen categories.
Approach
Bottom-up part-assembly representation: a Hypergraph Isomorphism Convolution makes part-feature fusion order- and repetition-invariant, and a leverage hypergraph (local-certainty class hyperedges plus global kNN hyperedges) with a memory-bank fuzzy reconstruction propagates embeddings from seen to unseen categories.
Key results
- On OP-COSEG, HAFR reaches 0.7015 mAP, about 8.7% above the second-best method HGM2R (0.6452).
- On OP-SHNP, mAP 0.5947 vs 0.5736 for HGM2R; ANMRR 0.2239 vs 0.2549.
- On OP-INTRA, mAP 0.5750 vs 0.5545 for HGM2R; ANMRR 0.4797 vs 0.4928.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HAFR rather than the strongest reported baseline | 3D object retrieval · OP-SHNP | mAP: 0.5736 -> 0.5947 | Helped |
| use HAFR rather than the strongest reported baseline | 3D object retrieval · OP-INTRA | mAP: 0.5545 -> 0.5750 | Helped |
| use HAFR rather than the strongest reported baseline | 3D object retrieval · OP-COSEG | mAP: 0.6452 -> 0.7015 | Helped |
| use HAFR rather than the strongest reported baseline | 3D object retrieval · OP-SHNP | NDCG: 0.5886 -> 0.5916 | Helped |
| use HAFR rather than the strongest reported baseline | 3D object retrieval · OP-INTRA | NDCG: 0.6019 -> 0.6382 | Helped |
| use HAFR rather than the strongest reported baseline | 3D object retrieval · OP-COSEG | NDCG: 0.7638 -> 0.7629 | Hurt |
| use HAFR rather than the strongest reported baseline | 3D object retrieval · OP-COSEG | ANMRR: 0.4355 -> 0.3604 | Helped |
| replace HIConv with GIN (ablation) | HAFR · OP-SHNP, OP-INTRA, and OP-COSEG | mAP: OP-SHNP: 0.5947 -> 0.5761; OP-INTRA: 0.5750 -> 0.5567; OP-COSEG: 0.7015 -> 0.6782 | Hurt |
| replace HIConv with MLP (ablation) | HAFR · OP-SHNP, OP-INTRA, and OP-COSEG | mAP: OP-SHNP: 0.5947 -> 0.5837; OP-INTRA: 0.5750 -> 0.5523; OP-COSEG: 0.7015 -> 0.6505 | Hurt |
| add Assembly Loss (Las) (ablation) | HAFR IAE · OP-SHNP, OP-INTRA, and OP-COSEG | mAP: OP-SHNP: 0.5767 -> 0.5947; OP-INTRA: 0.5560 -> 0.5750; OP-COSEG: 0.6829 -> 0.7015 | Helped |
| add Cross-Part Loss (Lxp) (ablation) | HAFR IAE · OP-SHNP, OP-INTRA, and OP-COSEG | mAP: OP-SHNP: 0.5767 -> 0.5947; OP-INTRA: 0.5561 -> 0.5750; OP-COSEG: 0.6828 -> 0.7015 | Helped |
| add the leverage hypergraph structure (ablation) | HAFR SFR · OP-SHNP, OP-INTRA, and OP-COSEG | mAP: OP-SHNP: 0.5755 -> 0.5947; OP-INTRA: 0.5545 -> 0.5750; OP-COSEG: 0.6887 -> 0.7015 | Helped |
Limitations
- Does not handle varying numbers of parts per object, as stated in the conclusion.
- Evaluated on only three datasets derived from ShapeNetPart, IntrA, and COSEG, with small training sets for OP-INTRA and OP-COSEG.
- Baselines are adapted close-set and open-set methods rather than methods designed for part-assembly open-set retrieval.
Experimental setup
Datasets: ShapeNetPart IntrA COSEG OP-SHNP OP-INTRA OP-COSEG
Benchmarks: OP-SHNP OP-INTRA OP-COSEG
Metrics: mAP NDCG ANMRR PR curve
Baselines: MMJN TCL SDML MMSAE PROSER HGM2R
Builds on: HGNN HGNNConv PROSER HGM2R
Closest papers by meaning
- Semi-Open 3D Object Retrieval via Hierarchical Equilibrium on Hypergraph (NeurIPS 2024 · similarity 0.89)
- Imagine: Image-Guided 3D Part Assembly with Structure Knowledge Graph (AAAI 2025 · similarity 0.86)
- 3DCoMPaT200: Language Grounded Large-Scale 3D Vision Dataset for Compositional Recognition (NeurIPS 2024 · similarity 0.86)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.84)
- Unsupervised 3D Structure Inference from Category-Specific Image Collections (CVPR 2024 · similarity 0.84)
- KP-RED: Exploiting Semantic Keypoints for Joint 3D Shape Retrieval and Deformation (CVPR 2024 · similarity 0.84)
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.