[D] Why I abandoned YOLO for safety critical plant/fungi identification. Closed-set classification is a silent failure mode

About this article

I’ve been building an open-sourced handheld device for field identification of edible and toxic plants wild plants, and fungi, running entirely on device. Early on I trained specialist YOLO models on iNaturalist research grade data and hit 94-96% accuracy across my target species. Felt great, until I discovered a problem I don’t see discussed enough on this sub. YOLO’s closed set architecture has no concept of “I don’t know.” Feed it an out of distribution image and it will confidently classi...