action recognition

**Action Recognition** is the **classification task of assigning a label to a trimmed video clip containing a single action** — serving as the foundational building block for more complex video understanding tasks. **What Is Action Recognition?** - **Definition**: Video Classification. Input: Video -> Output: Label (e.g., "Playing Tennis"). - **Constraint**: Usually assumes the video contains *only* the action of interest (trimmed). - **Benchmarks**: UCF101, HMDB51, Kinetics-400. **Why It Matters** - **Human-Computer Interaction**: Recognizing gestures to control devices (Kinect style). - **Content Moderation**: Automatically flagging violent or prohibited actions. - **Health**: Monitoring exercises or detecting falls in elderly care. **Architectures** - **Two-Stream**: Separate specialized networks for spatial (RGB frames) and temporal (Optical Flow) data. - **3D CNNs**: C3D, I3D (Inflated 3D ConvNets) that process time as a third dimension. - **Video Transformers**: ViViT, TimeSformer (applying attention across space and time). **Action Recognition** is **the "ImageNet Classification" of video** — the core capability that enables machines to identify human behavior.