Segment Anything Model (SAM)Meta AI 的可调prompt的图像分割基模型

4.8 (5)

审阅者 Daniel Nikulshyn·更新 2026年7月

Open Source Computer Vision Image Segmentation Foundation Model Research PyTorch Zero-Shot

概览

Segment Anything Model (SAM) 是由 Meta AI Research 开发的开源图像分割系统。只需提供图像和一个简单的提示（如一个点、一个框或一个粗略的掩码），它就能为几乎所有对象生成高质量的分割掩码，无需针对特定任务的训练。 SAM 在 SA-1B 数据集上进行训练，该数据集包含超过十亿个掩码，覆盖 1100 万张图像，使其具备强大的零样本泛化能力。它可以集成到计算机视觉流水线中，用于标注、图像编辑、医学影像、机器人、AR/VR 以及科学分析等任务。该模型和数据集采用宽松的授权条款发布，使得 SAM 成为研究者和开发者在无需从零训练定制模型的情况下进行灵活分割的常用构建模块。

主要功能

Promptable 分割使用点和盒子
自动掩膜生成整个图像
预训练的 ViT 基于图像编码器
零冲击转移到新领域
开源码和 SA-1B 数据集
集成 PyTorch 和常见的 CV 栈

价格

模型: Freemium
分类: 计算机视觉
评分: 4.8 / 5 (5)

使用场景

加速数据集标注

用SAM的可调prompt的分割来快速标注图像数据集中的对象，通过简单的点击或盒子减少手动注解时间，使ML训练管线更高效

图像编辑和叠加

生成精确的对象掩模用于背景去除、可选择编辑或叠加在创意工具中，避免训练定制分割模型

医学和科学成像分析

将零冲击分割应用于医学扫描或科学成像，对感兴趣的结构进行孤立，辅助测量和下游分析

机器人和 AR/VR 感知

将SAM整合到计算机视觉管线中，利用点或盒子提示，用于机器人操控或 AR/VR 场景理解中的对象孤立

优点 & 缺点

优点

在未见对象上具有强大的零冲击分割能力
灵活的提示：点、盒子或掩膜
开放的权重和大的公共数据集
易于集成通过官方 Python 库
零冲击转移到新领域
可用于各种场景
无需特定训练模型

缺点

大型模型可能对CPU实时操作有较大负载
未分配语义类标签
精度在极小或细结构上的下降
要求提示或自动掩膜生成设置

对决战绩

在万神殿中参与了 1 对决。

第1

第2

第3

Last battle

#2
Computer Vision Showdown — February 5, 2024
Feb 5, 2024 · #2 of 8

评测

4.8

5 个评分的平均值。

登录以留下评测。

Aaliyah Johnson

Apr 5, 2026

Years in this space

I've evaluated a lot of these over the years. What stands out here is pretrained ViT-based image encoder — handled better than most — and strong zero-shot segmentation on unseen objects. Worth the time if this is your use case.

Margaret Whitfield

Mar 5, 2026

Years in this space

I've evaluated a lot of these over the years. What stands out here is promptable segmentation with points and boxes — handled better than most — and flexible prompts: points, boxes, or masks. Worth the time if this is your use case.

Hannah Goldberg

Oct 15, 2025

Use it every day

Honestly didn't expect to like it this much. Integrates with PyTorch and common CV stacks is exactly what I needed, and easy to integrate via official Python library. I do wish requires prompts or automatic mask generation setup, but I reach for it almost every day now and it just clicks.

Olga Ivanova

Jul 10, 2025

Skeptical, then convinced

I went in skeptical — most tools in this space overpromise. It actually delivers on zero-shot transfer to new domains, and strong zero-shot segmentation on unseen objects caught me off guard. Requires prompts or automatic mask generation setup is why this isn't a perfect score, still, I'd recommend giving it a real trial.

Linda Petersen

May 29, 2025

Skeptical, then convinced

I went in skeptical — most tools in this space overpromise. It actually delivers on automatic mask generation for entire images, and open weights and large public dataset caught me off guard. still, I'd recommend giving it a real trial.

问答

Does SAM provide semantic class labels for the segmented objects?

No, SAM outputs binary masks only; it does not assign semantic class labels, so you would need an additional classifier if label information is required.

Asked by Giulia Conti · Mar 23, 2026

Can SAM be used for real‑time applications on CPU‑only hardware?

Because SAM’s ViT‑based encoder is large, it runs efficiently on GPUs but can be slow on CPUs, making real‑time CPU‑only deployment challenging for high‑throughput use cases.

Asked by Tariq Aziz · Feb 22, 2026

What input prompts does SAM accept for generating segmentation masks?

SAM can take point clicks, bounding boxes, or rough masks as prompts, and it also offers an automatic mode that generates masks for all objects in an image without any user input.

Asked by Salome Beridze · Dec 22, 2025