Large Vision-Language Models can precisely locate actions in videos without specialized training, using confidence scoring to identify key moments.

• No training data required: System works "out of the box" for new domains
• Confidence-based detection: Identifies precise action boundaries by measuring model certainty
• Cross-domain versatility: Effective for both medical and security applications
• Enhanced analysis: Enables detailed examination of specific motions in long videos

In medical settings, this technology transforms surgical video analysis by automatically identifying critical procedural steps, enabling better training and assessment without manual annotation.

Zero-shot Action Localization via the Confidence of Large Vision-Language Models