Fine-grained brand recognition in unstructured video faces high bounding box annotation costs and vulnerability to false positives. This article proposes and evaluates an annotation-efficient cascade architecture for Fine-Grained Visual Categorization (FGVC) in video. The approach couples a generalist detector (YOLO26) for product superclass localization with a specialist classifier (YOLO26-cls) for brand identification over the crops produced by the detector. We compared this cascade model against a single-stage detector under frame-by-frame validation, using 7,070 smartwatch images (Apple, Garmin, and Samsung) from 88 YouTube videos and five superclasses from the public Video E-commerce Retrieval Dataset (VERD). The cascade architecture reduced fine-grained dataset preparation time by approximately 22 times by eliminating manual bounding box annotation. While robustness to false positives proved to be a brand-dependent trade-off, the cascade maintained near-real-time processing throughput. Ultimately, the method's annotation efficiency and capacity for incremental expansion support its viability for video-based retail intelligence applications.