From Fixed Labels to Prompts: How Vision-Language Models Are Re-Wiring Object Detection

Object detection has evolved from hand-crafted features to deep CNNs with much higher accuracy, but most production systems are still stuck with fixed label sets that are expensive to update. New open-vocabulary, vision-language detectors (like Grounding DINO) let you detect arbitrary, prompt-defined concepts and achieve strong zero-shot performance on benchmarks, even without dataset-specific labels. The most practical approach today is hybrid: use these promptable models as teachers and auto-annotators, then distill their knowledge into small, closed-set detectors you can reliably deploy on edge devices.