98,000 Surveillance Images Annotated for Security AI
Challenge
The client was developing multiple AI models for security applications: person and vehicle detection across 38,000+ surveillance images, a 50,000-image person attribute dataset with gender and age classification, and a specialized head-covering detection system distinguishing between helmets, caps, masks, balaclavas, sunglasses, and other face-covering objects across 10,000 images.
Solution
We ran three parallel projects with up to 7 annotators working simultaneously. Starting with person/vehicle bounding boxes in YOLO format, we expanded to complex attribute annotation — labeling head and body regions with gender, age, and 9 head-covering categories. We built custom error-detection scripts to automate QA checks across the large datasets, catching annotation gaps before delivery.
Results
Successfully delivered across three distinct projects spanning person detection, attribute classification, and head-covering recognition. Quality improved measurably across iterations, and the client's marketing team returned a year later for additional collaboration — confirming long-term trust in our work.