SAM:可提示的分割一切模型Segment Anything
问题:分割任务通常按语义/实例/全景分开训练,换任务就要换模型;能不能定义一个「给提示就出掩码」的通用任务,用一个模型加提示工程覆盖多种分割?
证据位置:任务、模型与数据分别在 §3–§5;数据引擎与 SA-1B 在 §4;零样本评测在 §6–§7(23 个分割数据集、边缘检测、候选框、实例分割、文本到掩码);局限集中在 §8,结论紧随其后。
作者自陈局限:§8 列了六条——会漏细结构,偶尔幻觉出小的不连通部件,边界不如「放大再算」的方法锐利;给很多点时,专用交互式分割方法会更好;重编码器使整体并非实时(虽然提示解码是实时的);文本到掩码只是探索性尝试,不够稳健;如何用简单提示实现语义与全景分割仍不清楚;在特定领域,专用工具会优于 SAM。
它没主张的:没有说 SAM 在所有分割任务上超过专用模型——§8 明确说专用交互式方法与领域工具会更好;没有说文本到掩码可用——原文称 exploratory and not entirely robust;也没有说「分割一切」字面成立——论文自己把任务限定为可提示分割,并承认语义/全景分割的提示设计未解决。Problem: segmentation is usually split into semantic, instance and panoptic tasks with separate models; can a single “give a prompt, get a mask” task plus prompt engineering cover many segmentation settings?
Where the evidence sits: task, model and data in Sections 3-5; data engine and SA-1B in Section 4; zero-shot evaluation in Sections 6-7 (23 segmentation datasets, edge detection, object proposals, instance segmentation, text-to-mask); limitations in Section 8, immediately before the conclusion.
Stated limits: Section 8 lists six — SAM misses fine structures, occasionally hallucinates small disconnected components, and produces less crisp boundaries than methods that zoom in; dedicated interactive segmentation methods outperform it when many points are given; the heavy image encoder means overall performance is not real-time even though prompt decoding is; text-to-mask is exploratory and not robust; simple prompts for semantic and panoptic segmentation remain unclear; and domain-specific tools will beat SAM in their domains.
What it does not claim: it does not claim SAM beats specialised models on every segmentation task — Section 8 says dedicated interactive methods and domain tools do better; it does not claim text-to-mask is usable, calling it exploratory and not entirely robust; and it does not claim “segment anything” literally, framing the task as promptable segmentation and leaving semantic/panoptic prompting unresolved.
关键参数Key specs
- 证据位置 / evidence
- §3–5 任务/模型/数据;§6–7 零样本评测;§8 局限;结论
- 规模 / scale
- SA-1B:1B+ 掩码、1,100 万张授权与隐私合规图像
- 关键数字 / key numbers
- 23 个分割数据集零样本评测;单前景点即可产出高质量掩码
- 自陈局限 / stated limits
- 漏细结构、幻觉小部件、边界不锐利、整体非实时、文本到掩码不稳(§8)
- 复现状态 / reproduction
- 未复现
官方原文怎么说What the source says
over 1 billion masks on 11M licensed and privacy respecting images
It can miss fine structures, hallucinates small disconnected components at times, and does not produce boundaries as crisply as more computationally intensive methods
SAM’s overall performance is not real-time when using a heavy image encoder.
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-17
- 许可 / 类型Licence / kind
- arxiv-preprint
- 许可说明Licence evidence
- https://arxiv.org/abs/2304.02643
- 类目Category
- 架构Architecture
论文里的数字怎么读How to read the numbers
先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards