AdaptVPR: Route-Aware Hard Positive Generation for Robust Visual Place Recognition

AdaptVPR. Route-Aware Hard Positive Generation for Robust Visual Place Recognition. By Shunpeng Chen and eight others, from Beijing University of Posts and Telecommunications and Qilu University of Technology. Published on the third of September. Its code repository is public and real: about twenty-five source files.

Visual place recognition localizes a query image by retrieving database images of the same place or one nearby. Its robustness degrades under domain shift, and the list of shifts is the list of things that change about a place: illumination, weather, seasonal change, dynamic occlusion. One contributing factor sits upstream of the model entirely. Existing training data holds limited appearance diversity for the same place. If a place appears in the training set under one sky, no amount of architecture will tell a model what it looks like under another.

AdaptVPR is a route-aware generative augmentation framework that constructs same-place hard positives: images of the same place, made deliberately harder to match. The word route carries the design. Rather than applying one editing procedure everywhere, the framework decides per image which kind of change can safely be made, and only then makes it. The routes are complementary rather than alternatives, and the choice among them is taken image by image.

Deciding comes first. A vision language model parses scene attributes and estimates editing feasibility, and a rule-based scheduler determines the generation route from editability scores and risk constraints. Three complementary routes are available. The Global Appearance Route introduces global scene changes in weather, illumination and time of day. The Local Occlusion Route inserts plausible dynamic occluders. The Dual Route combines both kinds of perturbation to produce a more challenging shift. Every candidate is then evaluated by a verification scheme oriented to the task itself, based on geometric consistency and appearance diversity, which reduces the risk of structural drift while ensuring the appearance still varies enough to be useful. The routes differ in what happens on failure: Global candidates are generated once and rejected if verification fails, while Local Occlusion and Dual candidates use verification feedback for limited prompt refinement and regeneration.

Running the framework produces AdaptCities, one hundred and sixty thousand verified synthetic same-place hard positives. Experiments span multiple baselines and several vision foundation backbones rather than one favourable pairing, and report consistent gains on standard benchmarks together with substantial improvements under the challenging domain shifts that motivated the work. The R@1 gains reach up to nine point two percent. Larger gains under shift than on the standard benchmarks is the shape you would hope for, given what the augmentation was built to supply.

The caution in this design lives in the verification step. A generated image can look entirely plausible and still have drifted structurally, and the verification scheme exists to reduce exactly that risk before a candidate is accepted as a positive. Checking geometric consistency first is what keeps a hard positive from becoming a false one.

AdaptVPR. Route-Aware Hard Positive Generation for Robust Visual Place Recognition. By Shunpeng Chen and eight others, from Beijing University of Posts and Telecommunications and Qilu University of Technology. Published on the third of September. This is the long version: the whole paper rather than the abstract.

Visual place recognition turns localization into image retrieval: given a query image, find database images taken at the same place. It breaks under domain shift, because the same place under rain, at night, in snow, or behind traffic can look more different from itself than it does from somewhere else. The obvious remedy is more training data showing each place under more conditions, and collecting real revisits across weather, season and time of day is expensive and does not scale. Generative augmentation looks like the answer. The paper's argument is that generating images for this task is not the same problem as generating images.

Here is the difference. A useful generated image has to differ enough from its source to be an informative hard positive, while preserving the spatial structures that define the physical place. Optimising for realism, text alignment or editability does not give you that. A visually plausible rainy street or a realistic vehicle insertion may still alter building facades, road topology, lane markings, window layouts or scene boundaries, and those are exactly the cues place recognition depends on. The consequence is worse than a wasted sample: a structurally corrupted image treated as a positive during metric learning acts as a false positive, teaching the embedding space something untrue about which images show the same place.

Two gaps in existing pipelines follow from that. First, different domain shifts need different editing behaviour. Weather, illumination and time of day should change the scene globally while leaving spatial layout alone; vehicles and pedestrians should change localized regions without rewriting the background geometry. One static strategy handling both produces insufficient change, excessive structural modification, or implausible insertions. Second, existing pipelines generate candidates and then filter them, without feeding the failure back. So AdaptVPR puts a vision language model in front to parse scene attributes and estimate editing feasibility, and a rule-based scheduler picks the route from those editability scores under risk constraints. Three routes follow: a global appearance route, a local occlusion route, and a dual route combining both.

The cost of that loop is measured rather than waved at. A one-shot baseline with random route selection and no feedback takes a single attempt, about twenty-one seconds, and yields an acceptance rate near fifty-one percent. The full procedure takes on average a little over two attempts and roughly forty-six seconds, and its acceptance rate rises to about sixty-six percent. Running it produces AdaptCities, one hundred and sixty thousand verified synthetic same-place hard positives. Across multiple recognition methods and vision foundation backbones, the gains are consistent on standard benchmarks and substantially larger under challenging conditions, reaching over nine percent on top-one recall.

Boundaries are visible around the edges of the design. Verification is defined by geometric consistency and appearance diversity, which are proxies for place identity rather than place identity itself, so an edit that satisfies both while altering something recognition depends on would pass. The route menu is fixed at three, with a rule-based scheduler rather than a learned one, and the reflection loop runs within a budget, so difficult scenes converge or are abandoned. The generated shifts are also drawn from a familiar list, weather, illumination, time of day and dynamic occluders, which leaves whatever domain shift is not on that list unaddressed.

The transferable idea is that verification belongs inside the generation loop rather than after it. A filter placed at the end can only discard, and discarding tells you nothing about how to do better next time. A verifier consulted between rounds turns a failed candidate into an instruction, which is why the local and dual routes converge on images that a one-shot pipeline would simply have thrown away. What is being generated here is not more data. It is data that has survived a test designed around the task it will be used for.