fmv-grounding-detection

Closed-set object detection (GroundingDINO) for full-motion video and overhead imagery across land and maritime domains; detects a fixed ontology of six categories (person, ship, vehicle, aircraft, bike, maritime objects).

FMV-Grounding is a high-accuracy, closed-set object detection model for Full Motion Video (FMV) analytics that detects six mission-critical object classes: person, vehicle, bike, aircraft, ship, and maritime objects. The model is designed to operate robustly across horizontal, satellite, and aerial drone imagery, supporting a broad range of land and maritime operational scenarios. Built on a modern vision-language transformer architecture, it combines semantic understanding with multi-scale visual feature representation to accurately localize objects across varying viewpoints, resolutions, and environmental conditions. Language-guided object queries enable strong generalization across diverse geospatial scenes while maintaining precise detection performance for operationally important targets. It takes an image together with an optional text prompt naming any subset of its six supported categories, and returns bounding boxes with confidence scores. If no prompt is supplied, all six categories are detected. Class names outside this fixed set are not supported: they are ignored and produce no detections.