Research map / Vision-language and multimodal / Cross-modal learning
DistilVPR: Cross-Modal Knowledge Distillation 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 DistilVPR, a cross-modal relational knowledge distillation pipeline that lets single-modal VPR students learn from multi-modal teachers using self-agent and cross-agent relations across Euclidean, spherical, and hyperbolic manifolds.
Problem
Visual place recognition with multi-modal sensors improves accuracy but adds cost and is infeasible for lightweight systems; the goal is to transfer multi-modal teacher knowledge to single-modal students without extra sensors at inference.
Approach
Extends relational KD by computing relations both within agents (teacher-teacher, student-student) and across agents (teacher-student), and combines Euclidean, cosine (spherical), and Poincaré-ball hyperbolic distances to capture relationships across different curvatures.
Key results
- On Oxford RobotCar, DistilVPR-SC distills AdaFusion into AdaFusion-2D reaching 96.7 AR@1% / 89.0 AR@1, versus 94.2 / 84.2 for the student without distillation.
- On Boreas, DistilVPR-SC with MinkLoc++2D-Big teacher distills to MinkLoc++2D reaching 77.1 AR@1% / 65.4 AR@1, versus 75.2 / 60.0 without distillation.
- On Boreas 3D-to-2D (MinkLoc++3D teacher, MinkLoc++2D student), DistilVPR-SC reaches 81.1 AR@1% / 68.2 AR@1, the best among compared distillation methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine self-agent and cross-agent relationships (ablation) | DistilVPR · Boreas; MinkLoc++3D teacher → MinkLoc++2D | AR@1: 61.3 -> 68.2 | Helped |
| Use teacher-student/student-student relationships rather than teacher-teacher/teacher-student relationships (ablation) | DistilVPR · Boreas; MinkLoc++3D teacher → MinkLoc++2D | AR@1: 60.6 -> 66.4 | Helped |
| Add Euclidean, cosine, and hyperbolic relationship distances together (ablation) | DistilVPR-SC · Boreas; MinkLoc++3D teacher → MinkLoc++2D | AR@1: 66.8 -> 68.2 | Helped |
| Use Euclidean and cosine distances instead of Euclidean distance alone (ablation) | DistilVPR-SC · Boreas; MinkLoc++3D teacher → MinkLoc++2D | AR@1: 66.8 -> 67.4 | Helped |
| Use Euclidean and hyperbolic distances instead of Euclidean distance alone (ablation) | DistilVPR-SC · Boreas; MinkLoc++3D teacher → MinkLoc++2D | AR@1: 66.8 -> 66.8 | No clear effect |
| Use cosine and hyperbolic distances instead of cosine distance alone (ablation) | DistilVPR-SC · Boreas; MinkLoc++3D teacher → MinkLoc++2D | AR@1: 66.5 -> 66.9 | Helped |
| Distill with DistilVPR-SC rather than RKD | MinkLoc++2D student · Boreas; MinkLoc++3D teacher → MinkLoc++2D | AR@1: 60.4 -> 68.2 | Helped |
| Use a MinkLoc++3D teacher rather than a MinkLoc++ fusion teacher (ablation) | DistilVPR-SC with MinkLoc++2D student · Boreas | AR@1: 67.2 -> 68.2 | Helped |
| Use a MinkLoc++3D teacher rather than a MinkLoc++2D-big teacher (ablation) | DistilVPR-SC with MinkLoc++2D student · Boreas | AR@1: 65.4 -> 68.2 | Helped |
| Add a feature adaptor to the teacher rather than the student (ablation) | DistilVPR-SC · Boreas; MinkLoc++3D teacher → AdaFusion-2D | AR@1: 62.4 -> 63.1 | Helped |
Surprising findings
- A 3D-only teacher (MinkLoc++3D) gave better student distillation than the fusion teacher MinkLoc++, suggesting a good task performer is not necessarily a good teacher.
Limitations
- Assumes teacher and student have identical feature dimensions; a feature adaptor is needed otherwise (explored only in supplement).
- Gains on some settings are small or inconsistent; multi-manifold combinations of only two distances do not always help.
Experimental setup
Datasets: Oxford RobotCar Boreas
Benchmarks: Oxford RobotCar Boreas
Metrics: AR@1 AR@1%
Baselines: KD AFD EPC-Net RKD CSD LSD-Net MKD
Builds on: RKD CSD MinkLoc++ AdaFusion NetVLAD
Closest papers by meaning
- EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition (ICLR 2025 · similarity 0.88)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.87)
- CricaVPR: Cross-image Correlation-aware Representation Learning for Visual Place Recognition (CVPR 2024 · similarity 0.87)
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.86)
- SuperVLAD: Compact and Robust Image Descriptors for Visual Place Recognition (NeurIPS 2024 · similarity 0.86)
- Optimal Transport Aggregation for Visual Place Recognition (CVPR 2024 · similarity 0.86)
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.