temporal action localization

**Temporal Action Localization (TAL)** is a **video analysis task that predicts the start and end times of specific actions** — not just classifying *what* happened, but precisely pinpointing *when* it happened within an untrimmed video stream. **What Is Temporal Action Localization?** - **Input**: A long, untrimmed video (e.g., an hour of CCTV footage). - **Output**: A set of triplets ${Start, End, ClassLabel}$ for every action instance. - **Example**: "Run: 05:12-05:20", "Jump: 05:21-05:23". - **Metric**: mAP at different t-IoU (temporal Intersection over Union) thresholds. **Why It Matters** - **Video Editing**: Automatically creating highlight reels by finding exciting moments (e.g., goals in sports). - **Safety**: Detecting the exact moment a safety violation occurred in a factory. - **Efficiency**: Allows skipping hours of boring footage to find the 5 seconds of relevant activity. **Approaches** - **Proposal-based**: Generate candidate segments -> Classify them. - **Frame-level**: Classify every frame -> Group continuous positives. **Temporal Action Localization** is **the "Object Detection" of the time dimension** — drawing bounding boxes around time intervals instead of spatial regions.