FMV Grounding

FMV Grounding is a closed-set object detection model for full-motion video and overhead imagery that identifies six mission-critical categories — person, vehicle, bike, aircraft, ship, and maritime objects — across satellite, aerial, drone, and ground-based scenarios. Built on a vision-language transformer architecture, it pairs image understanding with language-guided object queries, and an optional text prompt can constrain detection to a specific subset of the supported classes — for example, only ship and maritime objects. Trained on roughly 3 million annotations and validated on 700,000, the model reaches 71.55% mAP and 83.43% average recall at IoU 0.5 on its internal test set.

Operational and geospatial intelligence workflows demand detectors tuned to a fixed, mission-relevant ontology rather than general-purpose recognition — objects outside the six categories are ignored by design, keeping outputs focused and auditable. On public benchmarks the model is especially strong on maritime and airborne targets, scoring 97.86% AP on Singapore Maritime ships and 98.76% on Mobdrone, with average inference latency near 200 ms on a single NVIDIA V100. Those characteristics suit maritime vessel monitoring, port and harbor analytics, drone-based perimeter surveillance, airfield detection, and border and coastal situational-awareness systems.