docs/en/datasets/semantic/ade20k.md
The ADE20K dataset is a large-scale semantic segmentation benchmark from MIT CSAIL with 20,210 training and 2,000 validation images densely annotated across 150 indoor, outdoor, object, and stuff categories. It is a standard resource for training and evaluating dense scene-understanding models with Ultralytics YOLO.
ADEChallengeData2016 archive and the Ultralytics ade20k.yaml config use only the training and validation splits.The Ultralytics configuration expects the official ADEChallengeData2016 layout:
ADEChallengeData2016/
├── images/
│ ├── training/
│ └── validation/
└── annotations/
├── training/
└── validation/
!!! warning "Manual Download Required"
ADE20K has no automatic download script. Download the ~1 GB [`ADEChallengeData2016.zip`](http://data.csail.mit.edu/places/ADEchallenge/ADEChallengeData2016.zip) archive and extract it directly into your `datasets/` folder. The archive's own top-level folder is already named `ADEChallengeData2016/`, so this produces `datasets/ADEChallengeData2016/` matching the layout above — do not create an `ADEChallengeData2016` folder yourself and extract into it, or you'll end up with a nested `datasets/ADEChallengeData2016/ADEChallengeData2016/` directory that the YAML won't find.
The masks_dir field is set to annotations, so each image under images/ is paired with its corresponding mask under annotations/. The original ADE20K masks use source label IDs where 0 is ignored, and the label_mapping section converts valid labels 1 through 150 to contiguous train IDs 0 through 149, mapping ignored pixels to 255.
ADE20K is widely used for training and evaluating deep learning models in semantic segmentation and scene parsing. Its diverse set of categories and complex scenes make it valuable for applications such as autonomous navigation, robotics, augmented reality, and image editing.
The breadth of indoor and outdoor scenes also makes ADE20K a strong benchmark for evaluating model generalization across domains. Pretrained YOLO26 semantic segmentation models reach up to 51.5 mIoU on the ADE20K validation set — see the semantic segmentation models page for the full benchmark table. ADE20K-format datasets are also fully compatible with Ultralytics Platform for dataset management and training.
A dataset YAML file defines the ADE20K paths, classes, mask directory, and label mapping. The ade20k.yaml file is maintained at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/datasets/ade20k.yaml.
!!! example "ultralytics/cfg/datasets/ade20k.yaml"
```yaml
--8<-- "ultralytics/cfg/datasets/ade20k.yaml"
```
To train a YOLO26n-sem model on the ADE20K dataset for 100 epochs with an image size of 512, you can use the following code snippets. For a comprehensive list of available arguments, refer to the model Training page.
!!! example "Train Example"
=== "Python"
```python
from ultralytics import YOLO
# Load a model
model = YOLO("yolo26n-sem.pt") # load a pretrained model (recommended for training)
# Train the model
results = model.train(data="ade20k.yaml", epochs=100, imgsz=512)
```
=== "CLI"
```bash
# Start training from a pretrained *.pt model
yolo semantic train data=ade20k.yaml model=yolo26n-sem.pt epochs=100 imgsz=512
```
ADE20K images are released for non-commercial research and educational use only; the dataset's annotation software is separately licensed under BSD-3. Commercial use requires permission from MIT CSAIL.
If you use the ADE20K dataset in your research or development work, please cite the following paper:
!!! quote ""
=== "BibTeX"
```bibtex
@inproceedings{zhou2017scene,
title={Scene Parsing through ADE20K Dataset},
author={Zhou, Bolei and Zhao, Hang and Puig, Xavier and Fidler, Sanja and Barriuso, Adela and Torralba, Antonio},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
year={2017}
}
```
We would like to acknowledge the MIT CSAIL Computer Vision Group for creating and maintaining this valuable resource for the computer vision community. For more information about the ADE20K dataset and its creators, visit the ADE20K dataset website.
The ADE20K dataset is a large-scale scene parsing benchmark used for semantic segmentation, with 20,210 training and 2,000 validation images publicly released across 150 categories covering indoor, outdoor, object, and stuff classes. Researchers use ADE20K because of its diverse scenes, fine-grained category set, and standardized evaluation metrics like mean Intersection over Union (mIoU), which make it ideal for benchmarking dense prediction models.
To train a YOLO26n-sem model on the ADE20K dataset for 100 epochs with an image size of 512, you can use the following code snippets. For a detailed list of available arguments, refer to the model Training page.
!!! example "Train Example"
=== "Python"
```python
from ultralytics import YOLO
# Load a model
model = YOLO("yolo26n-sem.pt") # load a pretrained model (recommended for training)
# Train the model
results = model.train(data="ade20k.yaml", epochs=100, imgsz=512)
```
=== "CLI"
```bash
# Start training from a pretrained *.pt model
yolo semantic train data=ade20k.yaml model=yolo26n-sem.pt epochs=100 imgsz=512
```
The ADE20K dataset follows the official ADEChallengeData2016 layout, with images organized under images/training/ and images/validation/, and corresponding masks under annotations/training/ and annotations/validation/. The Ultralytics YAML file pairs each image with its mask via the masks_dir: annotations field, and uses label_mapping to convert source label IDs 1–150 into contiguous train IDs 0–149, mapping the ignore label to 255.
Yes. Download the ADEChallengeData2016.zip archive (~1 GB) and extract it directly into your datasets/ folder before training — the archive's own top-level folder is already named ADEChallengeData2016/, so extracting it there (not into a separate ADEChallengeData2016 folder you create yourself) produces the images/ and annotations/ layout that ade20k.yaml expects.
label_mapping?ADE20K annotation masks store source label IDs where 0 denotes the ignore or background class. The label_mapping section maps valid labels 1 through 150 to contiguous train IDs 0 through 149, and assigns 255 to ignored pixels so they are excluded from the loss and metrics during training and validation.
No. ADE20K images are released under terms restricting use to non-commercial research and education; the accompanying annotation software is separately licensed under BSD-3. Contact MIT CSAIL for commercial licensing options.