Research map / Label-efficient and robust learning / Recognition and long-tail
SuperVLAD: Compact and Robust Image Descriptors for Visual Place Recognition
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 SuperVLAD, a NetVLAD variant without cluster centers and with few clusters (4), plus a 1-Cluster VLAD using a ghost cluster, for compact and domain-robust VPR global descriptors on DINOv2 features.
Problem
Global descriptors for visual place recognition from VLAD-style aggregation are high-dimensional, and learned cluster centers become improper under train-test domain shift, degrading generalization.
Approach
Remove the learned cluster centers from NetVLAD so that soft-assigned features are aggregated by weighted sum of raw local features, use only a few clusters for compactness, and add a ghost cluster that is discarded from the output to yield a 1-cluster descriptor of GeM-level dimension.
Key results
- DINOv2-based SuperVLAD reaches 95.0 R@1 on Pitts30k and 93.2 R@1 on SPED, outperforming SALAD and CricaVPR with a 3072-dim descriptor.
- On cross-domain test (trained on Pitts30k, tested on MSLS-val), SuperVLAD with CCT improves R@1 by 6.8 points over NetVLAD (55.4 to 62.2).
- 1-Cluster VLAD (768-dim) achieves 90.4 R@1 on MSLS-val versus 85.4 for GeM pooling and 88.4 for class token with DINOv2-base.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace NetVLAD with SuperVLAD by removing cluster centers (ablation) | DINOv2-B VPR · MSLS-val, trained on Pitts30k | R@1: 72.7 -> 75.1 | Helped |
| replace NetVLAD with SuperVLAD by removing cluster centers (ablation) | CCT VPR · MSLS-val, trained on Pitts30k | R@1: 55.4 -> 62.2 | Helped |
| add a ghost cluster (ablation) | CCT-SV trained on Pitts30k · MSLS-val | R@1: 60.3 -> 62.2 | Helped |
| add a ghost cluster (ablation) | DINOv2-SV trained on GSV-Cities · MSLS-val | R@1: 92.4 -> 92.2 | Hurt |
| use 1-Cluster VLAD instead of GeM pooling (ablation) | DINOv2-B global descriptor · Pitts30k-test | R@1: 89.5 -> 91.6 | Helped |
| use 1-Cluster VLAD instead of the class token (ablation) | DINOv2-B global descriptor · MSLS-val | R@1: 88.4 -> 90.4 | Helped |
| add the cluster-wise cross-image encoder (ablation) | SuperVLAD with 4 clusters · Pitts30k-test | R@1: 92.6 -> 95.0 | Helped |
| add the cluster-wise cross-image encoder (ablation) | SuperVLAD with 4 clusters · MSLS-val | R@1: 92.2 -> 92.2 | No clear effect |
| use 16 or 64 clusters instead of 4 (ablation) | SuperVLAD without cross-image encoder · MSLS-val | R@1: 92.2 -> 90.5; 92.2 -> 91.4 | Hurt |
| increase cluster count from 4 to 16 or 64 (ablation) | SuperVLAD without cross-image encoder · Pitts30k-test | R@1: 92.6; 92.4; 92.5 | No clear effect |
| remove cluster centers and use SuperVLAD rather than NetVLAD (ablation) | CCT VPR · Pitts30k-test, trained on Pitts30k | R@1: 84.7 -> 84.7 | No clear effect |
| remove cluster centers and use SuperVLAD rather than NetVLAD (ablation) | ViT VPR · Pitts30k-test, trained on MSLS | R@1: 79.0 -> 79.4 | Helped |
Surprising findings
- With a CNN (VGG16) backbone, SuperVLAD with 64 clusters underperforms NetVLAD, so the advantage depends on transformer backbones.
- With DINOv2 backbone trained on GSV-Cities, the ghost cluster makes almost no difference, and removing it slightly helps on some metrics.
Limitations
- Relies on a transformer backbone; with CNN backbones and many clusters it is inferior to NetVLAD.
- The cross-image encoder requires multi-image inference batches and is ineffective with single-image queries, so it is kept optional.
Experimental setup
Datasets: GSV-Cities MSLS Pitts30k
Benchmarks: Pitts30k-test MSLS-val Nordland SPED-test Pitts250k Tokyo247 Baidu Mall
Metrics: Recall@1 Recall@5 Recall@10 inference time
Baselines: NetVLAD SFRS TransVPR CosPlace MixVPR EigenPlaces SelaVPR CricaVPR SALAD GeM pooling class token
Builds on: NetVLAD GhostVLAD CricaVPR GeM pooling DINOv2
Closest papers by meaning
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.91)
- Optimal Transport Aggregation for Visual Place Recognition (CVPR 2024 · similarity 0.91)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.88)
- MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering (NeurIPS 2025 · similarity 0.88)
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.87)
- EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition (ICLR 2025 · similarity 0.87)
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.