ProVision is a video captioning agent built for Track 2. It analyzes short video clips and returns captions that match the requested tone, including formal, sarcastic, humorous technical, and humorous non-technical styles. The system focuses on describing visible subjects, actions, setting, and mood clearly while avoiding unnecessary assumptions. It is designed to generalize to unseen clips, produce concise JSON output for automated judging, and stay within the Docker-based evaluation workflow. For each task, ProVision samples representative visual moments from the video, interprets what is happening, and generates polished style-specific captions that preserve factual accuracy while matching the target writing style. The goal is simple: reliable captions that are useful, readable, and aligned with the evaluator format.
Category tags: