LVIS Dataset

The LVIS dataset is a large-scale object detection and instance segmentation benchmark with 1,203 object categories across roughly 160,000 images. Developed and released by Facebook AI Research (FAIR), it reuses the COCO images but adds a much larger, long-tailed vocabulary — from common objects like cars and bicycles to rarer ones such as handbags, umbrellas, and sports equipment — to push progress on recognizing infrequent categories.

<strong>Watch:</strong> YOLO World training workflow with LVIS dataset

Key Features

• LVIS spans roughly 160,000 images with about 2 million instance annotations for object detection and instance segmentation.
• The dataset defines 1,203 object categories, including common objects like cars, bicycles, and animals as well as fine-grained categories such as umbrellas, handbags, and sports equipment.
• Annotations include object bounding boxes and segmentation masks, with a strong focus on rare, long-tailed categories.
• LVIS provides standardized evaluation metrics like mean Average Precision (mAP) for detection and mean Average Recall (mAR) for segmentation, making model comparison straightforward.
• LVIS uses the same images as the COCO dataset but with different splits and far more detailed annotations.

Dataset Structure

The Ultralytics lvis.yaml configuration defines three splits:

The upstream LVIS benchmark also includes a held-out test set of roughly 20,000 images whose ground-truth annotations are not public; results are submitted to the LVIS evaluation server for scoring.

Applications

The LVIS dataset is widely used to train and evaluate deep learning models for object detection (such as YOLO, Faster R-CNN, and SSD) and instance segmentation (such as Mask R-CNN). Its large, long-tailed vocabulary, high annotation volume, and standardized evaluation metrics make it an essential benchmark for measuring how well computer vision models handle rare categories.

To label your own images, train, and manage large-vocabulary datasets like LVIS in your browser, run the full workflow with Ultralytics Platform.

Dataset YAML

The lvis.yaml file defines the dataset configuration — the dataset paths, class names, and other metadata. It is maintained in the Ultralytics repository at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/cfg/datasets/lvis.yaml.

!!! example "ultralytics/cfg/datasets/lvis.yaml"

```yaml
--8<-- "ultralytics/cfg/datasets/lvis.yaml"
```

Usage

To train a YOLO26n model on the LVIS dataset for 100 epochs with an image size of 640, you can use the following code snippets. The dataset (20.1 GB) downloads automatically on first use. 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.pt")  # load a pretrained model (recommended for training)

    # Train the model
    results = model.train(data="lvis.yaml", epochs=100, imgsz=640)
    ```

=== "CLI"

    ```bash
    # Start training from a pretrained *.pt model
    yolo detect train data=lvis.yaml model=yolo26n.pt epochs=100 imgsz=640
    ```

Sample Images and Annotations

The LVIS dataset contains diverse images with many object categories in complex scenes. Below is an example of a mosaiced training batch:

• Mosaiced Image: This training batch combines multiple LVIS images into one through mosaicing, a technique that increases the variety of objects and scenes in each batch and helps the model generalize to different object sizes, aspect ratios, and contexts.

Citations and Acknowledgments

If you use the LVIS dataset in your research or development work, please cite the following paper:

!!! quote ""

=== "BibTeX"

    ```bibtex
    @inproceedings{gupta2019lvis,
      title={LVIS: A Dataset for Large Vocabulary Instance Segmentation},
      author={Gupta, Agrim and Dollar, Piotr and Girshick, Ross},
      booktitle={Proceedings of the {IEEE} Conference on Computer Vision and Pattern Recognition},
      year={2019}
    }
    ```

We would like to acknowledge the LVIS Consortium for creating and maintaining this valuable resource for the computer vision community. For more information about the LVIS dataset and its creators, visit the LVIS dataset website.

FAQ

What is the LVIS dataset used for?

The LVIS dataset is used to train and benchmark object detection and instance segmentation models on a large, long-tailed vocabulary. Developed by Facebook AI Research (FAIR), LVIS features 1,203 object categories and about 2 million instance annotations, which makes it an essential resource for measuring how well models like Ultralytics YOLO recognize both common and rare categories.

How many images and classes are in the LVIS dataset?

The Ultralytics lvis.yaml configuration covers 1,203 object categories split into 100,170 training images, 19,809 validation images, and a 5,000-image minival subset that is identical to the COCO val2017 set. The images are the same as those in COCO, but LVIS annotates them with a far larger vocabulary.

How can I train a YOLO26n model using the LVIS dataset?

To train a YOLO26n model on the LVIS dataset for 100 epochs with an image size of 640, follow the example below.

!!! example "Train Example"

=== "Python"

    ```python
    from ultralytics import YOLO

    # Load a model
    model = YOLO("yolo26n.pt")  # load a pretrained model (recommended for training)

    # Train the model
    results = model.train(data="lvis.yaml", epochs=100, imgsz=640)
    ```

=== "CLI"

    ```bash
    # Start training from a pretrained *.pt model
    yolo detect train data=lvis.yaml model=yolo26n.pt epochs=100 imgsz=640
    ```

For detailed training configurations, refer to the Training documentation.

How does the LVIS dataset differ from the COCO dataset?

The images in the LVIS dataset are the same as those in the COCO dataset, but the two differ in their splits and annotations. LVIS provides a larger and more detailed vocabulary with 1,203 object categories compared to COCO's 80 categories. LVIS also focuses on annotation completeness and diversity, aiming to push the limits of object detection and instance segmentation models by offering more nuanced and comprehensive data.

Does the LVIS dataset include a test set?

The Ultralytics lvis.yaml configuration provides train (100,170 images), validation (19,809 images), and minival (5,000 images) splits. The upstream LVIS benchmark also has a held-out test set of roughly 20,000 images whose ground-truth annotations are not public — results are submitted to the LVIS evaluation server for scoring rather than evaluated locally.

Why should I use Ultralytics YOLO for training on the LVIS dataset?

Ultralytics YOLO models, including the latest YOLO26, are optimized for real-time object detection with state-of-the-art accuracy and speed. They support a wide range of annotations, such as the fine-grained ones provided by the LVIS dataset, making them ideal for advanced computer vision applications. Ultralytics also offers seamless integration with training, validation, and prediction modes, ensuring efficient model development and deployment.