We introduce the Segment Anything (SA) project: a new task, model, and dataset for image segmentation. Using our efficient model in a data collection loop, we built the largest segmentation dataset to date (by far), with over 1 billion masks on 11M licensed and privacy respecting images. The model is designed and trained to be promptable, so it can transfer zero-shot to new image distributions and tasks. We evaluate its capabilities on numerous tasks and find that its zero-shot performance is impressive -- often competitive with or even superior to prior fully supervised results. We are releasing the Segment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and 11M images at https://segment-anything.com to foster research into foundation models for computer vision.
核心贡献 · Key contributions
提出了可提示分割任务，实现了对新图像分布和任务的零样本泛化。 Introduced the promptable segmentation task, enabling zero-shot generalization to new image distributions and tasks.
开发了分割一切模型（SAM），具有灵活的提示编码器和快速掩码解码器，支持实时交互使用。 Developed the Segment Anything Model (SAM) with a flexible prompt encoder and fast mask decoder for real-time interactive use.
构建了 SA-1B 数据集，包含 1100 万张图像上的超过 10 亿个掩码，是最大的分割数据集。 Built the SA-1B dataset, the largest segmentation dataset with over 1 billion masks on 11 million images.
在 23 个多样化分割数据集上展示了强大的零样本性能，通常与全监督方法相竞争。 Demonstrated strong zero-shot performance on 23 diverse segmentation datasets, often competitive with fully supervised methods.
以开放许可发布了 SAM 和 SA-1B，以促进计算机视觉基础模型的研究。 Released SAM and SA-1B under open license to foster research into foundation models for computer vision.
通过分摊图像嵌入，在 CPU 上实现了约 50 毫秒的实时掩码预测。 Achieved real-time mask prediction from prompts in ~50ms on CPU via amortized image embedding.
局限 · Limitations
SAM 可能遗漏精细结构并产生小的不连续组件幻觉。 SAM may miss fine structures and hallucinate small disconnected components.
与计算密集型的放大方法相比，边界不够清晰。 Boundaries are less crisp compared to computationally intensive zoom-in methods.
当提供多个点时，专用交互式分割方法优于 SAM。 Dedicated interactive segmentation methods outperform SAM when many points are provided.
文本到掩码任务尚处于探索阶段，不够稳健。 Text-to-mask task is exploratory and not entirely robust.
使用重型图像编码器时，整体性能无法达到实时。 Overall performance is not real-time when using a heavy image encoder.
论文章节 · Sections（共 15）
摘要Abstract
1 引言1 Introduction
2 分割一切任务2 Segment Anything Task
3 分割一切模型3 Segment Anything Model
4 分割一切数据引擎4 Segment Anything Data Engine
5 分割一切数据集5 Segment Anything Dataset
6 分割一切 RAI 分析6 Segment Anything RAI Analysis
7 零样本迁移实验7 Zero-Shot Transfer Experiments
7.1 零样本单点有效掩码评估7.1 Zero-Shot Single Point Valid Mask Evaluation