How to play
Select the class that matches an image region. Paint the complete area without crossing its visible boundary. Evaluate your work for pixel, overlap, boundary, and efficiency scores.
Keyboard numbers select palette classes. Brackets change brush size. Control or Command plus Z restores the previous mask state.
Segmentation concepts
Semantic segmentation assigns a class to every pixel. Instance segmentation also separates different objects sharing one class. Panoptic segmentation combines both tasks.
Intersection over Union measures overlap between predicted and correct masks. Dice score gives another overlap measure. Boundary accuracy focuses on object edges.
Real-world uses
Segmentation helps autonomous vehicles understand roads and people. Medical systems outline organs or suspicious regions. Mapping systems classify land, water, crops, and buildings.
Balanced classes matter because large backgrounds can hide poor object performance. Inspect per-class scores instead of relying only on total accuracy.
Questions and guidance
What is the difference between a mask and an image?
An image stores visible colours and textures. A mask stores class labels for pixels. This game displays masks using class-specific colours.
Why can pixel accuracy look high when objects are missed?
A large background may dominate the image. Mean IoU gives each class more balanced importance. Review recall for smaller objects.
How does the smart brush work?
The smart brush stays inside the truth region touched at stroke start. It demonstrates edge-aware assistance while still requiring class selection.
What does the error heatmap show?
Green pixels were classified correctly. Red pixels were wrong or left unpainted. Strong red boundaries reveal localisation problems.