SA-1B
Dataset
SA-1B (Segment Anything 1-Billion) is a large-scale image segmentation dataset released by Meta, containing 11 million high-resolution images and over 1 billion automatically generated high-quality segmentation masks for training general object segmentation models.
Dataset Highlights
Ultra-large scale image segmentation dataset, promoting the development of general visual segmentation models
Ultra-large Scale
Contains over 1.1 billion segmentation masks, making it the largest image segmentation dataset to date, over 400 times larger than the previous largest open segmentation dataset.
High Resolution
The average image resolution is 3300x4950 pixels, with all images having a short side of no less than 1500 pixels, ensuring high-quality segmentation annotations.
SAM Auto-generated
Masks are fully automatically generated by the Segment Anything Model (SAM), undergoing multiple rounds of quality filtering and validation, with extremely high annotation consistency and accuracy.
Diverse Scenes
Covers various scenes including indoor, outdoor, natural, and urban, with a rich variety of object categories, supporting the construction of truly general segmentation models.
Privacy Protection
All faces in the images have been blurred, complying with privacy protection standards, and can be safely used for academic research and commercial development.
Multi-granularity Segmentation
From whole objects to fine parts, masks cover multiple granularity levels, supporting various segmentation tasks from coarse to fine granularity.
Applicable Scenarios
Covers a wide range of computer vision tasks from academic research to industrial applications
Image Segmentation
Train and evaluate various image segmentation models such as semantic segmentation, instance segmentation, and panoptic segmentation
Zero-Shot Transfer
Utilize the SA-1B pre-trained model for zero-shot transfer to new domains and new categories of segmentation tasks
Base Model Training
Provide large-scale high-quality training data for visual base models, promoting the development of general visual intelligence
Medical Image Analysis
Transfer learning to medical imaging segmentation scenarios, assisting in lesion area detection, organ segmentation, and other clinical applications
Data Preview
The following is an example of mask annotation corresponding to a single image (JSON format)
{
"image": {
"image_id": 4,
"width": 2250,
"height": 1500,
"file_name": "sa_4.jpg"
},
"annotations": [
{
"id": 1,
"segmentation": {
"size": [1500, 2250],
"counts": "eW`a05kc06J..."
},
"area": 38279,
"bbox": [174, 263, 243, 338],
"predicted_iou": 0.9621,
"stability_score": 0.9724,
"point_coords": [[295.5, 432.0]],
"crop_box": [0, 0, 2250, 1500]
}
]
}
3 Steps to Get Started Quickly
Quickly start your image segmentation project from browsing to usage
Browse Dataset
View the details of the SA-1B dataset on the Ace Data Cloud platform, including metadata such as data scale, annotation format, and licensing agreements.
Download Data
Download images and corresponding mask annotation files on demand, with data provided in chunks, allowing flexible selection of the required subset.
Use with SAM
Use the segment_anything library to load the pre-trained SAM model, fine-tuning or evaluating it with the SA-1B data.
Start Exploring the SA-1B Dataset
A large-scale image segmentation dataset with open licensing, available immediately. Whether you are a computer vision researcher or an industrial developer, SA-1B is an ideal choice for training general segmentation models.