Research map / Architectures and efficiency / Attention and transformers
BoQ: A Place is Worth a Bag of Learnable Queries
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 Bag-of-Queries (BoQ), a transformer aggregation head with learnable global queries that cross-attend to backbone features for one-stage visual place recognition, outperforming SOTA global methods and two-stage reranking methods.
Problem
Visual place recognition needs compact, robust global descriptors under viewpoint, seasonal, and illumination changes; Transformer-based VPR has relied on reranking and self-attention-derived queries, with weaker one-stage global retrieval than CNN methods.
Approach
Replace input-dependent queries with a fixed bag of learnable global queries that self-attend among themselves and then cross-attend to local features from the backbone, with outputs from multiple BoQ blocks concatenated and projected into an L2-normalized descriptor.
Key results
- BoQ with ResNet-50 (16384 dim) reaches 91.4% R@1 on MSLS-val and 74.4% R@1 on Nordland*, outperforming MixVPR and EigenPlaces.
- On SVOX Night, BoQ achieves 85.2% R@1, 20.8 points above the second best method (MixVPR).
- BoQ without reranking (91.4% R@1 on MSLS-val) surpasses two-stage methods Patch-NetVLAD, TransVPR and R2Former, and is over 30x faster than R2Former.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the number of learnable queries from 8 to 64 per BoQ block (ablation) | BoQ · MSLS-val | Recall@1: 88.1 -> 91.3 | Helped |
| Increase the number of learnable queries from 8 to 64 per BoQ block (ablation) | BoQ · AmsterTime | Recall@1: 44.3 -> 52.0 | Helped |
| Use 4 BoQ blocks instead of 1 (ablation) | BoQ with ResNet-18 · MSLS | Recall@1: 87.1 -> 88.1 | Helped |
| Increase BoQ blocks from 4 to 8 (ablation) | BoQ with ResNet-18 · MSLS | Recall@1: 88.1 -> 87.2 | Hurt |
| Replace ResNet-50 with ResNet-101 backbone (ablation) | BoQ · MSLS-val | Recall@1: 91.4 -> 90.5 | Hurt |
| Replace ResNet-50 with ResNet-101 backbone (ablation) | BoQ · Nordland* | Recall@1: 74.4 -> 71.9 | Hurt |
| Add self-attention between the learned queries (ablation) | BoQ with ResNet-18 · Nordland | Recall@1: 56.4 -> 65.9 | Helped |
| Add self-attention between the learned queries (ablation) | BoQ with ResNet-18 · MSLS-val | Recall@1: 87.1 -> 88.4 | Helped |
| Use BoQ instead of MixVPR | Visual place recognition retrieval · Nordland* | Recall@1: 58.4 -> 74.4 | Helped |
| Use BoQ instead of EigenPlaces | Visual place recognition retrieval · MSLS-val | Recall@1: 89.2 -> 91.4 | Helped |
| Use BoQ instead of EigenPlaces | Visual place recognition retrieval · Pitts30k | Recall@1: 92.5 -> 92.4 | Hurt |
| Use BoQ instead of MixVPR | Visual place recognition retrieval · SVOX Night | Recall@1: 64.4 -> 85.2 | Helped |
Surprising findings
- Using ResNet-101 did not outperform ResNet-50, which the authors attribute possibly to memory-forced smaller batch sizes.
- BoQ with a ResNet-18 backbone remains competitive with ResNet-50-based CosPlace and MixVPR.
Limitations
- Reranking with spatial information from the last encoder output is left as future work.
- Performance on AmsterTime and some multi-view benchmarks lags behind the best reported results (e.g., EigenPlaces on Pitts30k).
Experimental setup
Datasets: GSV-Cities
Benchmarks: MSLS-val Pitts250k-test Pitts30k-test AmsterTime Eynsham Nordland SPED St Lucia SVOX
Metrics: Recall@1 Recall@5 Recall@10
Baselines: AVG GeM NetVLAD SPE-NetVLAD Gated NetVLAD Conv-AP CosPlace MixVPR EigenPlaces Patch-NetVLAD TransVPR R2Former
Builds on: NetVLAD MixVPR DETR Multi-Similarity loss
Closest papers by meaning
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.87)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.86)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.86)
- Optimal Transport Aggregation for Visual Place Recognition (CVPR 2024 · similarity 0.86)
- DistilVPR: Cross-Modal Knowledge Distillation for Visual Place Recognition (AAAI 2024 · similarity 0.85)
- EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition (ICLR 2025 · similarity 0.85)
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.