video question answering

**Video Question Answering (VideoQA)** is a **multimodal task where a model answers natural language questions about a video clip** — requiring the integration of visual features, temporal dynamics, and linguistic semantics to understand events that unfold over time. **What Is VideoQA?** - **Definition**: Given Video $V$ and Question $Q$, predict Answer $A$. - **Complexity**: Unlike ImageQA, the answer often depends on *when* something happens or the *sequence* of actions. - **Types**: - **Descriptive**: "What is the man doing?" - **Temporal**: "What did he do after opening the door?" - **Causal**: "Why did the car stop?" **Why It Matters** - **Search**: "Find the part of the meeting where we discussed the budget." - **Surveillance**: "Did anyone enter the room between 2 PM and 3 PM?" - **Accessibility**: Helping visually impaired users understand dynamic content. **Key Challenges** - **Long-Term Dependency**: Remembering details from the start of a long video to answer a question at the end. - **Multimodal Fusion**: aligning audio (speech), visual frames, and motion information. **Video Question Answering** is **comprehension for dynamic scenes** — testing an AI's ability to maintain a coherent mental model of a changing world.