Locality Aware Visual Analysis

LAVA presents a rapid training process of predictable duration for semantically enhanced object detectors that can quickly produce a detector with predictable performance. Current production systems usually require weeks or months to produce and maintain an object detector, with most of the time spent in data curation, and incorporate analyst feedback post-production. LAVA accelerates detector development process by introducing analyst-in-the-loop guided iterative feedback loops throughout the detector production process. As one of its inputs, LAVA accepts analyst feedback in the form of text prompts and uses foundational language vision models to perform context-aware few-shot or zero-shot object detection and classification. LAVA uses semantic segmentation and contextual information such as land cover maps and street maps to reduce the false alarm rate while maintaining a high recall rate for all objects, including rare objects in novelcontexts.