VDC-Agent: When Video Detailed Captioners Evolve Themselves via Agentic Self-Reflection

ABSTRACT

Generating highly detailed, temporally accurate video captions requires models to understand complex spatial and temporal dynamics. VDC-Agent introduces an autonomous framework where the captioning model iteratively critiques and refines its own output using an agentic self-reflection loop. By cross-referencing generated text against specific video frames autonomously, the system resolves hallucinations and produces state-of-the-art dense video descriptions.