The GeoFMV Podcast

Why did we build GeoFMV when tools like the QGIS FMV plugin and ArcGIS Pro's Image Analyst extension already exist? We walk through the pain points — expensive licenses, Qt5 tech debt, no local AI — and explain how GeoFMV addresses them without breaking the analyst's workflow.

A technical breakdown of the binary metadata format powering every military drone video. We cover the KLV encoding model (Key-Length-Value), the 16-byte MISB ST 0601 Universal Key, BER length encoding, the 146 defined tags, and why the round-trip property matters for data integrity.

We compare running YOLOv11 locally via ONNX Runtime against routing frames to OpenRouter's multimodal vision models. Topics: air-gap requirements, GPU acceleration (DirectML/CUDA), NMS post-processing, the confidence threshold tradeoff, and how both modes feed the same georeferencing pipeline.

The ArcGIS Pro Image Analyst extension costs thousands per seat — and most organisations only need FMV capabilities for a handful of analysts. We explain how LibVLCSharp replaces the native VideoLayer, how a custom C# BinaryReader parses MISB KLV without any proprietary SDK, and how OpenCvSharp4 handles frame extraction for AI inference.

DuckDB WASM ships an entire analytical database engine to the browser as WebAssembly. We walk through the spatial, httpfs, and zipfs extensions that power GeoFMV's analytics layer — enabling analysts to query detection archives from S3, join with PostGIS geodatabases, and explore 3D Tiles metadata, all without a server round-trip.

Most DJI drones output raw MP4 video and a separate CSV flight log — not STANAG 4609. We walk through the GeoFMV multiplexor that ingests a DJI video + CSV pair, maps DJI columns to MISB ST 0601 tags, interpolates timestamp gaps, and produces a STANAG 4609 MPEG-TS file that can be exploited with all the same tools as native FMV.

In a contested or classified environment, sending video frames to a cloud API is not an option. We cover how to set up GeoFMV's local AI mode: selecting and packaging ONNX models, configuring GPU execution providers (DirectML on Windows, CUDA on Linux), managing model updates on air-gapped networks, and the custom ONNX model workflow for domain-specific detection classes.

AI detections are proposals, not verdicts. We walk through the full analyst review workflow: the detection review panel, accept/reject/reclassify actions, map symbology by review status, the click-to-seek integration between the detection layer and the video player, and export filtering by status, class label, confidence, and time range into GeoPackage, GeoJSON, and MISB ST 0903 VMTI format.