Zero-shot means asking a model to do something for which it saw no specific training examples — recognise a new object, classify a new category — and getting a usable answer because it learned broad concepts. It trades peak accuracy for instant coverage of new cases.
The book · Volume 6 of 7
AI for Video: A Complete Guide to Intelligence in Video Products: From Models and Inference to Computer Vision, Speech, Search, and Generative Video
Zero-shot models let a product recognise new objects or categories without collecting training data, at the cost of lower peak accuracy than a tuned model. Volume 6 covers zero-shot detection and classification in detail within computer vision: open-vocabulary detection, embeddings, model artifacts and the ML minimum for deciding when zero-shot is good enough. Written by Nikolay Sapunov, CEO at Fora Soft.
Can a zero-shot model cover your detection needs without custom training?
Send us sample footage and the classes you care about, and we will test zero-shot models on it and tell you where custom training is still needed. 250+ video products shipped since 2005.
