EVENT DOSSIER
The AdaptVPR framework was developed to construct a synthetic dataset containing 160,000 pieces of data.
To address the issue of reduced robustness in visual place recognition (VPR) caused by domain shift, Shunpeng Chen et al. proposed the AdaptVPR framework. This approach utilizes a visual language model to parse scene attributes and evaluate edit feasibility, and combines a rule-based scheduler to generate three complementary types of difficult positive samples: global appearance routes (changing weather, lighting, and time), local occlusion routes (inserting dynamic occlusions), and dual routes (combining both). The generated candidate images are filtered through geometric consistency and appearance diversity; if failed, optimization is provided. Based on this, the AdaptCities dataset containing 160,000 verified synthetic data was constructed. Experiments show that this method achieves performance improvements in multiple benchmark tests and challenging domain shift scenarios, with a maximum R@1 gain of 9.2%. The related code and data are available publicly.
政府 · 科研机构across 1 days
The AdaptVPR framework was developed to construct a synthetic dataset containing 160,000 pieces of data.
Shunpeng Chen et al. proposed the AdaptVPR framework, using a visual language model and a rule-based scheduler to generate difficult positive samples for three complementary routes (global appearance, local occlusion, dual routes), and constructed a collection containing 160,000 validated synthetic data pieces.
2 reportsShunpeng Chen 等人提出 AdaptVPR，一种基于路线感知的生成增强框架，旨在提升视觉地点识别在挑战性外观变化下的鲁棒性。该框架通过三条场景依赖路线生成同地点困难正样本：全局外观路线（改变天气、光照和时间）、局部遮挡路线（插入动态遮挡物）以及双路线（结合两者）。研究构建包含 16 万条已验证合成数据的 AdaptCities 数据集，并采用平衡几何一致性与外观多样性的验证方案配合反馈驱动提示优化。实验表明，AdaptVPR 在多种视觉地点识别基线及视觉基础骨干网络上均能持续提升检索性能。
Visual Place Recognition (VPR) robustness often declines due to domain offset. AdaptVPR proposes a generative enhancement framework for route perception to address the issue of insufficient visual diversity at the same location. This framework utilizes a visual language model to parse scene attributes and evaluate editing feasibility, combining a rule-based scheduler to determine the generated route. The process is divided into three complementary modes: global appearance, local occlusion, and dual routes. Generated candidate images are screened using a VPR validation scheme based on geometric consistency and appearance diversity; if the validation fails, global candidates are rejected or local/dual route candidates are suggested for optimization. Based on this, the AdaptCities dataset containing 160K validated synthetic data was constructed. Experiments show that this method achieves consistent improvement in multiple benchmark tests and under challenging domain offset conditions, with a maximum R@1 gain of 9.2%. The relevant code and data are available publicly.