A new benchmark wants to know if AI can tell two nearly identical noodle shops apart from ten meters away.
Researchers released MMS-VPR, a dataset for visual place recognition built inside a roughly 70,800-square-meter pedestrian district in Chengdu, China. It contains 110,529 images and 2,527 video clips spanning 208 fine-grained location classes, some just 10-20 meters apart. The team paired fresh 2024 field footage, shot from multiple angles and at different times of day, with seven years of social media photos from 2019 to 2025, totaling 31,726 geolocated images tagged with GPS data, timestamps, and text descriptions. Alongside the dataset, they built MMS-VPRlib, a benchmarking platform that ran 22 existing models, ranging from shallow classifiers to transformers and graph neural networks, against the data.
Most place-recognition datasets come from car-mounted cameras cruising wide roads in Western cities, which suits self-driving cars but not a delivery robot or AR app trying to navigate a crowded Chengdu shopping street where every storefront looks the same. This dataset targets that exact gap, and the numbers back up the premise: combining video, text, and image data pushed accuracy to 98.1%, a 19.2-point jump over using photos alone, with video's temporal information doing most of the heavy lifting.
That gulf between recognizing a highway on-ramp and recognizing which identical bubble tea stand you are standing in front of is exactly where most consumer navigation and augmented-reality apps still stumble.