Research map / Label-efficient and robust learning / Recognition and long-tail
QI-IRA: Quantum-Inspired Interactive Ranking Aggregation for Person Re-identification
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 QI-IRA, a quantum-inspired interactive ranking aggregation method that updates fusion weights of basic re-ID rankings from relevance feedback on top-K items, matching or beating fully-supervised RA with as few as two feedbacks per query.
Problem
Ranking aggregation of multiple person re-ID rankings is either unsupervised (no external supervision to optimize) or fully-supervised (costly labeling), so a label-efficient method that adapts fusion weights is needed.
Approach
Models gallery images as quantum particles and basic re-ID methods as observation operators, interpreting rank aggregation as linear weighting of density operators (via Gleason's theorem); fusion weights are updated from the ranking scores of user-feedback positive and negative samples with a single momentum parameter gamma.
Key results
- On Market1501, MARS and DukeMTMC-VideoReID, QI-IRA with K=1 and T=2 (two feedbacks per query) outperforms fully-supervised RA methods in mAP.
- At K=5 and T=3 on DukeMTMC-VideoReID, QI-IRA reaches 99.36% mAP, about 2.15% above IRAR.
- QI-IRA(1,1) is roughly 0.7 ms aggregation time vs. 6-9 ms for fully-supervised methods, with interaction cost of about 3 s per gallery item; QI-IRA is less time-costly than fully-supervised RA on CUHK03 (detected).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QI-IRA with two feedback samples per query (K=2, T=1) rather than the unsupervised RA baselines | QI-IRA · Market1501, DukeMTMC-reID, CUHK03 detected and labeled, MARS, DukeMTMC-VideoReID | mAP | Helped |
| Use two feedback samples per query instead of one (K=2 rather than K=1; T=1) (ablation) | QI-IRA · Market1501 | mAP: 94.55 -> 94.74 | Helped |
| Use two feedback samples per query instead of one (K=2 rather than K=1; T=1) (ablation) | QI-IRA · DukeMTMC-reID | mAP: 88.63 -> 88.74 | Helped |
| Use two feedback samples per query instead of one (K=2 rather than K=1; T=1) (ablation) | QI-IRA · CUHK03 detected | mAP: 84.41 -> 84.54 | Helped |
| Use two feedback samples per query instead of one (K=2 rather than K=1; T=1) (ablation) | QI-IRA · CUHK03 labeled | mAP: 86.93 -> 87.26 | Helped |
| Use two feedback samples per query instead of one (K=2 rather than K=1; T=1) (ablation) | QI-IRA · MARS | mAP: 87.15 -> 87.72 | Helped |
| Use two feedback samples per query instead of one (K=2 rather than K=1; T=1) (ablation) | QI-IRA · DukeMTMC-VideoReID | mAP: 95.81 -> 97.01 | Helped |
| Increase feedback samples per round and interaction rounds (ablation) | QI-IRA · Market1501, DukeMTMC-reID, CUHK03 detected and labeled, MARS, DukeMTMC-VideoReID | mAP | Helped |
| Set the weight-update fusion factor γ between 0.5 and 0.9 (ablation) | QI-IRA · Six re-ID datasets | mAP | No clear effect |
| Use QI-IRA with only two rounds of interaction at K=1 rather than the fully-supervised RA methods | QI-IRA · Market1501, MARS, DukeMTMC-VideoReID | mAP | Helped |
Surprising findings
- The initial mean-based aggregation alone does not exceed the best fully-supervised method; gains come from the feedback rounds.
- Sensitivity to the momentum parameter gamma is low: results stay stable for gamma between 0.5 and 0.9.
Limitations
- Relies on simulated or real user relevance feedback; the paper's time estimates assume 3 seconds per gallery interaction.
- Performance is not best on DukeMTMC-reID in mAP at K=2 because basic rankings have more items, limiting gains when K x T is small.
- The theoretical motivation is a quantum analogy; the paper does not validate it beyond empirical results.
Experimental setup
Datasets: Market1501 DukeMTMC-reID CUHK03 (detected) CUHK03 (labeled) MARS DukeMTMC-VideoReID
Benchmarks: Market1501 DukeMTMC-reID CUHK03 (detected) CUHK03 (labeled) MARS DukeMTMC-VideoReID
Metrics: Rank@1 mAP
Baselines: CombMIN CombMAX CombSUM CombANZ CombMNZ BordaCount Dowdall Median RRF iRANK Mean HPA
Builds on: IRA CSRA Mean RA Gleason's theorem
Closest papers by meaning
- Re-Ranking Person Re-Identification With k-Reciprocal Encoding (CVPR 2017 · similarity 0.84)
- Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification (CVPR 2025 · similarity 0.83)
- Re-Ranking via Metric Fusion for Object Retrieval and Person Re-Identification (CVPR 2019 · similarity 0.82)
- Learning to Attack Real-World Models for Person Re-identification via Virtual-Guided Meta-Learning (AAAI 2021 · similarity 0.82)
- Human-centered Interactive Learning via MLLMs for Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.81)
- Bipartite Ranking From Multiple Labels: On Loss Versus Label Aggregation (ICML 2025 · similarity 0.81)
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.