DAPR Framework Cuts Geographic Bias in VPR by 18%
Long-tail data causes VPR failure in less-visited areas; DAPR rebalances gradients.
Urban-scale Visual Place Recognition (VPR) systems suffer from a hidden long-tail problem: frequently photographed locations dominate training data, causing models to fail in sparsely covered areas. A new paper, accepted to ECCV 2026, systematically characterizes this geographic imbalance and proposes Distribution-Aware Place Recognition (DAPR), a model-agnostic plug-in framework. DAPR rebalances gradient contributions across head and tail classes during training and introduces a multi-scale distance search mechanism at retrieval to compute per-class distributional compactness.
Tested on the large-scale SF-XL benchmark, DAPR outperforms the prior classification-retrieval baseline by 18.3% on test set v1 and 6.7% on test set v2. As a plug-in, it also delivers consistent improvements across representative VPR methods on SF-XL, MSLS, and Pitts30k, proving its generalizability. The work highlights how geographic bias in urban datasets undermines VPR reliability and offers a practical, method-agnostic solution for robust place recognition in real-world deployments.
- DAPR is model-agnostic and rebalances gradient contributions from head and tail geographic classes.
- Achieves 18.3% improvement on SF-XL test set v1 and 6.7% on test set v2 over previous baselines.
- Consistently boosts performance across VPR methods on SF-XL, MSLS, and Pitts30k benchmarks.
Why It Matters
Makes VPR reliable in less-photographed urban areas, critical for autonomous navigation and mapping.