RGB images with corresponding pixel-wise annotations. Images are containing realistic everyday scenes.
For evaluation the usual metric is mean Intersection over Union (mIoU).
Available data is split into train and val sets, containing 1464 and 1449 images respectively.
- 20 class labels + background + unlabeled
- Median number of pixels: 187500, 90% of images have between 147000 and 200000 pixels
- Median height-to-width ratio is 0.75, 90% of images have between 0.66 and 1.5.
Links to helper scripts for extraction/transformation/visualization.
| Link | Description |
|---|---|
| Dataset scripts | A collection of utility functions in Python |