Best Zero-Shot Object Detection Models

These models ground words in appearance, so concrete visual phrases beat category jargon: "orange striped shirt" outperforms "non-compliant uniform", and "person wearing a hard hat" beats "PPE compliance". Attribute plus noun is the sweet spot; long relational descriptions ("the shirt next to the woman holding a coffee") degrade into phrase-grounding territory where precision drops. When a concept keeps failing, try synonyms: detectors often know "high-visibility vest" but not your industry term for it.

Score distributions vary wildly between prompts on the same model: 0.3 can be conservative for "person" and hopelessly noisy for "cracked weld". Treat the threshold as per-prompt configuration, not a global constant, and set it by inspecting a handful of your own images rather than trusting defaults.

Zero-shot output looks convincing, so measure it: label a small ground-truth set and check both misses and false positives before wiring the detector into anything automated. It shines for prototyping, ad hoc analysis, and pre-labeling datasets; for a fixed class list at production volume, use it to bootstrap labels and train a specialized detector, which will beat it on accuracy, speed, and cost. The open-vocabulary object detection task page covers choosing between the detectors themselves.