Researchers have released YILDIZ-VPR, a new dataset for teaching computers to recognize where a photo was taken outdoors.
The team at Yildiz Technical University built it by walking the same routes across their Davutpasa campus over and over, recording video with a GoPro 9 camera synced to GPS. The footage covers historical buildings, modern structures, roads, green areas, and wooded paths, shot at different times of day, across seasons, and in varying weather. Each extracted frame is tagged not just with GPS coordinates but also gyroscope, speed, and temperature readings from onboard sensors. The dataset targets visual place recognition, the task of matching a query photo to its location by comparing it against a bank of geo-tagged images.
Most visual place recognition datasets get collected once, under one set of conditions, then models are benchmarked on best-case lighting. YILDIZ-VPR's repeated walks over an extended period mean the same corner of campus shows up in daylight, dusk, rain, and snow, which is closer to what a delivery robot or navigation app actually has to handle. That kind of long-term visual variability, more than sheer data volume, is what tends to separate localization models that work in a lab from ones that hold up outside it.
It is a niche academic contribution, not a product launch, and the real test is whether other labs actually adopt a single-campus dataset as a shared benchmark rather than filing it away.