I geolocated an animal abuse video using OSINT

Описание: This video walks through a text-based semantic search methodology for narrowing a large geolocation search space.

The case study involves identifying the exact location of a short video taken somewhere in London. First, OS Open Roads dataset was used to identify the ~80k candidate T-junctions. Next, OSM was used to filter to the ~20k junctions on 2-lane roads, and Google street view panos pulled for these. Lastly, CLIP / SigLIP models were used to score against natural-language prompts describing visible features in the footage.

A two-pass architecture was used. ViT-B/16 on all 20k candidates, then SigLIP SO400M to refine the top decile with additional directional crops (e.g., railings on both sides of a crossing)

The workflow could be improved by (1) adding image-to-image matching alongside text prompts where a reference view of the target scene is available, (2) incorporating late-interaction multi-vector scoring for features that occupy small portions of the frame, and (3) offloading the 2nd rescoring run to cloud GPU compute (e.g., Lambda Labs, RunPod, Modal) to enable faster prompt iteration and larger candidate sets.

The back half of the video covers the follow-through after the geolocation was solved; reaching out to nearby vet practices, the RCVS Code of Professional Conduct disclosure exception (Ch 14), and the current state of responses from the RSPCA and the vet practices.

This video is intended for educational purposes, focusing on methodology and analytical reasoning using publicly available data.