LVIS Dataset

v1.0

Training set

• 1,270,141 instances (1 GB)

• 100,170 images (18 GB)

Validation set

• 244,707 instances (192 MB)

• 19,809 images (1 GB)

Test Dev

• info (4 MB)

• 19,822 images (6 GB)

Test Challenge

• info (4 MB)

• 19,822 images (6 GB)

Please see our recommended best practices for using this data.

Note: LVIS uses the COCO 2017 train, validation, and test image sets. If you have already downloaded the COCO images, you only need to download the LVIS annotations. LVIS val set contains images from COCO 2017 train in addition to the COCO 2017 val split.

Data Format

LVIS has annotations for instance segmentations in a format similar to COCO. The annotations are stored using JSON. The LVIS API can be used to access and manipulate annotations. The JSON file has the following format:

 {  
 info : info
 images : [images],
 annotations: [annotations],
 licenses : [licenses],
 }  

 info{  
 year : int
 version : str,
 description  : str,
 contributor  : str,
 url  : str,
 date_created  : datetime,
 }  

 license{  
 id  : int
 name  : str,
 url  : str,
 }  

We describe the data structures a bit more in detail below.

Images

Each image now comes with two additional fields.

• not_exhaustive_category_ids : List of category ids which don't have all of their instances marked exhaustively.

• neg_category_ids : List of category ids which were verified as not present in the image.

• coco_url : Image URL. The last two path elements identify the split in the COCO dataset and the file name (e.g., http://images.cocodataset.org/train2017/000000391895.jpg). This information can be used to load the correct image from your downloaded copy of the COCO dataset.

 image{  
id  : int
width  : int,
height  : int,
license  : int,
flickr_url : str,
coco_url  : str,
date_captured  : datetime,
not_exhaustive_category_ids  : [int],
neg_category_ids  : [int],
 }  

Categories

LVIS categories are loosely based on WordNet synsets.

• synset : Provides a unique string identifier for each category. Loosely based on WordNet synets.

• synonyms : List of object names that belong to the same synset.

• def : The meaning of the synset. Most of the meanings are derived from WordNet.

• image_count : Number of images in which the category is annotated.

• instance_count : Number of annotated instances of the category.

• frequency : We divide the categories into three buckets based on image_count in the train set.

 categories{  
id  : int
synset  : str,
synonyms  : [str],
def  : str,
instance_count  : int,
image_count  : int,
frequency  : str,
 }  

Annotations

The segmentation format in LVIS is always a list of polygons.

 annotation{  
id  : int
image_id  : int,
category_id  : int,
segmentation  : [polygon],
area  : float,
bbox : [x,y,w,h],
 }  

License

The LVIS annotations along with this website are licensed under a Creative Commons Attribution 4.0 License. All LVIS dataset images come from the COCO dataset; please see link for their terms of use.