qhjqhj00 long-term-vpr-eval
Long-Term Visual Localization in Dynamic Benthic Environments: A Dataset, Footprint-Based Ground Truth, and Visual Place Recognition Benchmark — Larsen et al. (2026) (arXiv:2603.04056, 2026)
What this evaluates
Evaluates long-term visual place recognition (VPR) capabilities in dynamic underwater benthic environments. It probes a model's ability to geolocate camera views over multi-year intervals despite habitat changes, varying terrain ruggedness, and sub-decimeter registration errors.
Datasets
- Benthic Reference Sites Dataset — total ?; splits: benchmark (-1)
Metrics
Recall@K(primary) — range: percent- Recall@K measures the fraction of query images for which at least one ground-truth matching database image appears in the top-K retrieved results. Ground-truth matches are defined by a footprint-based IoU threshold of τ_f ≥ 0.07.
Input / output format
Input: Stereo camera images (query and database views) from different survey visits over benthic sites, along with camera pose/calibration data for footprint computation.
Output: Ranked list of database images retrieved for each query image.
Scoring recipe
def compute_recall_at_k(queries, database, top_k, iou_thresh=0.07):
hits = 0
for q in queries:
retrieved = get_top_k_retrievals(q, database, top_k)
if any(compute_footprint_iou(q, db) >= iou_thresh for db in retrieved):
hits += 1
return (hits / len(queries)) * 100
Common pitfalls
- Using traditional location-based ground truth overestimates VPR performance in rugged terrain; the paper validates footprint-based matching as superior.
- Ignoring residual registration errors between multi-year visits can cause false positives; a conservative footprint IoU threshold (τ_f ≈ 0.07) is required.
- Performance is highly dependent on seafloor type and revisit interval, so averaging across all sites without stratification hides critical model weaknesses.
Evidence (verbatim from paper)
To assess how the selected VPR models generalize across seafloor types, we evaluate their mean Recall@K across all visit pairs for each reference site. This provides insight into how seafloor characteristics influence localization performance. As shown in Figure 7, we report the mean Recall@K over a range of rank cutoffs K, illustrating how retrieval accuracy varies with the number of candidates considered. Ground-truth correspondences are defined using footprint-based camera links established through the methodology in Section 2.5.2, applying the lower footprint IoU threshold (τ_f=0.07) derived in Section 3.2 to compensate for residual registration errors between visits.
Citation
@misc{larsen2026longterm,
title={Long-Term Visual Localization in Dynamic Benthic Environments: A Dataset, Footprint-Based Ground Truth, and Visual Place Recognition Benchmark},
author={Larsen et al. (2026)},
year={2026},
note={arXiv:2603.04056}
}
- arXiv: 2603.04056