Research map / Image generation / Autoregressive and latent generation
LOCORE: Image Re-ranking with Long-Context Sequence Modeling
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 LOCORE, a list-wise image re-ranker that jointly processes a query and up to 100 gallery images at the local-descriptor level using a Longformer sequence model with token-level binary classification and sliding-window inference.
Problem
Image re-ranking in instance-level retrieval typically scores each query-gallery pair independently with local descriptors, ignoring interactions among gallery images; list-wise re-ranking has only been done with global descriptors.
Approach
Concatenate query and gallery local descriptors into one long sequence with [SEP] tokens, process it with a pre-trained Longformer using global attention on query and [SEP] tokens, train with per-token binary cross-entropy under shuffled gallery order, and extend list size at test time with a sliding window.
Key results
- On RPar+1M hard with DELG RN50 descriptors, LOCORE-base reaches 50.5 mAP vs 46.1 for CVNet Reranker (3,072 descriptors).
- On CUB-200, LOCORE-base improves R@1 by +9.4 and mAP@R by +15.0 over the global baseline, while RRT gets -0.2 R@1 and +5.8 mAP@R.
- LOCORE-small has about 3x lower latency than LOCORE-base-comparable pair-wise methods and can re-rank 100 images in a single forward pass (18.5 to 140.1 ms latency across variants).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LOCORE-base instead of CVNet Reranker | DELG RN50 re-ranking · ROxf+1M Hard | mAP: 46.1 -> 50.5 | Helped |
| use LOCORE-base instead of CVNet Reranker | DELG RN50 re-ranking · RPar+1M Hard | mAP: 52.5 -> 58.2 | Helped |
| use LOCORE-base instead of global descriptors | image retrieval · CUB-200 | R@1: 68.9 -> 78.3 | Helped |
| remove gallery shuffled training (ablation) | LOCORE-tiny · SOP | R@1: 82.4 -> 80.7 | Hurt |
| remove global query attention (ablation) | LOCORE-tiny · SOP | R@1: 82.4 -> 60.7 | Hurt |
| shuffle gallery candidates at evaluation (ablation) | LOCORE-tiny · SOP | R@1: 82.4 -> 81.8 ±0.1 | Hurt |
| use first token score aggregation (ablation) | LOCORE-base · SOP | R@1: 83.8 -> 83.7 | No clear effect |
| use [SEP] token score aggregation (ablation) | LOCORE-base · SOP | mAP@R: 71.0 -> 70.9 | No clear effect |
| increase sliding-window re-ranking from N=100 to N=200, stride S=50 (ablation) | LOCORE-small · ROxf+1M Medium | mAP: 82.7 -> 83.9 | Helped |
| increase sliding-window re-ranking from N=100 to N=200, stride S=50 (ablation) | LOCORE-small · RPar+1M Medium | mAP: 84.0 -> 85.1 | Helped |
| increase model size from LOCORE-tiny to LOCORE-base (ablation) | LOCORE · SOP | mAP@R: 68.0 -> 71.0 | Helped |
Surprising findings
- Scaling up pair-wise re-rankers such as RRT does not yield notable gains, whereas LOCORE improves consistently with model size.
- Without global query attention, LOCORE suffers severe performance drop (mAP 53.0 vs 65.1 global baseline), and without gallery shuffled training it merely repeats global retrieval results.
Limitations
- Sliding window inference is needed to exceed the training list size, adding runtime cost for large shortlists.
- Requires local descriptors extracted from a backbone (DELG or DINOv2) and a list of globally retrieved candidates, so it depends on the first-stage retriever.
Experimental setup
Datasets: Google Landmarks v2 (clean) CUB-200 Stanford Online Products In-Shop
Benchmarks: Revisited Oxford (ROxf) Revisited Paris (RPar) ROxf+1M RPar+1M CUB-200 Stanford Online Products In-Shop
Metrics: mAP Recall@k mAP@R latency FLOPs peak memory
Baselines: RRT CVNet Reranker AMES GV SuperGlobal Rerank AQE DQE αQE SSR Rerank
Builds on: Longformer RRT AMES CVNet DELG
Closest papers by meaning
- INQUIRE: A Natural World Text-to-Image Retrieval Benchmark (NeurIPS 2024 · similarity 0.85)
- Very Efficient Listwise Multimodal Reranking for Long Documents (ICML 2026 · similarity 0.84)
- MagicLens: Self-Supervised Image Retrieval with Open-Ended Instructions (ICML 2024 · similarity 0.84)
- Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers (ICLR 2025 · similarity 0.84)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.84)
- FashionERN: Enhance-and-Refine Network for Composed Fashion Image Retrieval (AAAI 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.