video object segmentation

**Video Object Segmentation (VOS)** is a **dense prediction task that assigns a pixel-level mask to specific objects in every frame of a video** — requiring the model to propagate a segmentation mask from the first frame (semi-supervised) or automatically discover primary objects (unsupervised). **What Is VOS?** - **Task**: Tracking the exact shape (mask) of an object as it moves and deforms. - **Semi-Supervised (One-Shot)**: User provides mask for Frame 1 -> Model segments Frames 2 to $N$. - **Zero-Shot**: Model automatically segments the most salient moving object. **Why It Matters** - **VFX & Rotoscoping**: Automating the tedious task of cutting out actors for special effects. - **Video Editing**: "Change the color of this car" throughout the entire video clip. - **Robotics**: Precise manipulation of moving objects (catching a ball). **Challenges**: - **Occlusion**: Object gets covered -> Re-appears later (Re-identification). - **Drift**: Small errors in frame 2 accumulate to massive errors by frame 100. **Video Object Segmentation** is **the "Green Screen" of AI** — capable of digitally extracting moving subjects from their environment without a studio setup.