spatiotemporal detection

**Spatiotemporal Detection** (or Video Object Detection) is the **task of tracking and classifying objects across both space and time** — essentially drawing a "tube" (sequence of bounding boxes) around an object as it moves through a video. **What Is Spatiotemporal Detection?** - **Goal**: Detect objects in every frame and link them consistently. - **Output**: A 3D volume (Tube) in the $H imes W imes T$ space. - **Challenge**: Motion blur, occlusion (object disappears behind a tree), and deformation. **Why It Matters** - **Autonomous Driving**: Tracking pedestrians and cars is not a per-frame task; the system needs to know "Target ID 42 is moving left". - **Sports Analytics**: Tracking a specific player or the ball throughout a match. - **Behavior Analysis**: Understanding interactions (e.g., "Person A handed an object to Person B"). **Key Datasets** - **AVA (Atomic Visual Actions)**: Detects actions localized in space and time. - **ImageNet VID**: Object detection in video. **Spatiotemporal Detection** is **4D perception** — understanding that objects are continuous entities that persist through time, not just flickering pixels in isolated frames.