Breaking01 / AI
Lens tests satellite search without target-specific training
Lens searches image features with reduced positional bias and reports performance at the best threshold for each land-cover class.
The authors of Lens reported on 16 Nov 2026 that they evaluated satellite-image search on a test set without training a detector for the target class; the preprint is not peer-reviewed and does not report independent replication.
The paper describes a method for finding similar objects and scenes using a frozen image model, whose parameters remain unchanged. After reducing positional bias in its image features, the method measures similarity between features around a selected point and those in the imagery being searched.
The authors reported a mean pixel-level F1 score, a measure balancing precision and recall, of 0.39 (n=5,128 tiles, DFC2020 test set, best threshold per class). This result uses the best decision threshold selected separately for each land-cover class.
The paper also reports applying the search tool to satellite imagery of agricultural, industrial and desert areas. Compared with INSID3, a pairwise comparison method used as a baseline, Lens recorded higher search scores across multiple land-cover classes and lower query costs.
The abstract does not report results from independent researchers replicating the evaluation, leaving independent replication status unspecified.