分割一切模型SAM的潜力与展望:综述

Potential and prospects of segment anything model:a survey

随着基于对比文本—图像对的预训练(contrastive language-image pre-training,CLIP)方法或者模型、聊天生成预训练转换器(chat generative pre-trained Transformer,ChatGPT)、生成预训练转换器-4(generative pre-trained Transformer-4,GPT-4)等基础大模型的出现,通用人工智能(artificial general intelligence, AGI)的研究得到快速发展。AGI旨在为人工智能系统赋予更强大的执行能力,使其能够自主学习、不断进化,解决各种问题和处理不同的任务,从而在多个领域得到广泛应用。这些基础模型在大规模数据集上进行训练后,能够成功应对多样的下游任务。在这一背景下,Meta公司提出的分割一切模型(segment anything model,SAM)于2023年取得重要突破,在图像分割领域获得了优异的性能,以至于被称为图像分割终结者。其原因之一是,通过SAM数据引擎方法用三阶段采集的、包含1 100万图像和超过10亿掩码的分割一切—十亿(segment anything 1 billion,SA-1B)图像分割数据集,同时保证了掩码的品质和多样性,继续导致在分割领域的突破。在SAM开源后不久,科研人员提出了一系列改进的方法和应用。为了能全面深入了解分割一切模型的发展脉络、优势与不足,本文对SAM的研究进展进行了梳理和综述。首先,从基础模型、数据引擎和数据集等多个方面简要介绍了分割一切模型的背景和核心框架。在此基础上,本文详细梳理了目前分割一切模型的改进方法,包括提高推理速度和增进预测精度两个关键方向。然后,深入探讨分割一切模型在图像处理任务、视频相关任务以及其他领域中的广泛应用。这一部分详细介绍了模型在各种任务和数据类型上的卓越性能,突出其在多个领域的泛用性和发展潜力。最后,对分割一切模型未来的发展方向和潜在应用前景进行了深入分析和讨论。

The emergence of foundational large-scale models,such as contrastive language-image pre-training(CLIP),chat generative pre-trained Transformer(ChatGPT),and generative pre-trained Transformer-4(GPT-4),has facilitated thesignificant growth of the field of artificial general intelligence(AGI).AGI aims to imbue systems with the ability to performvarious tasks,which enables them to learn autonomously and evolve.This broad applicability spans various domains and is intended to address diverse problems and accomplish numerous downstream tasks.These models,after being trained onmassive datasets,possess the capability to handle a multitude of downstream tasks.In this context,Meta’s segment any⁃thing model(SAM)has substantially progressed and introduced the largest image segmentation dataset to date,that is,SA1B.This dataset includes over 11 million images and more than one billion mask in 2023.One reason is that SA-1B wascollected through SAM’s data engine approach in three stages.This approach simultaneously ensures the quality and diver⁃sity of these masks,which contributes significantly to breakthroughs in the segmentation domain.This development hasprofoundly impacted the advancements in the foundational models in the field of computer vision.This study provides acomprehensive understanding of the SAM framework through a detailed review and analysis of relevant research.First,thisstudy delves into three aspects of the background and basic framework of the SAM model.The first aspect involves the tasksof SAM,including traditional image segmentation and prompt-guided interactive image segmentation.The second aspect isthe model architecture of SAM,encompassing image encoders,prompt encoders,and mask decoders.The third aspectrevolves around the data,including the data engine for collecting datasets and dataset SA-1B.Building upon this founda⁃tion,the study then organizes and analyzes methods for improving the SAM model from two perspectives.The first perspec⁃tive is enhancing inference speed.The reason is that improved inference speed reduces the deployment costs of SAM,which makes it more convenient for application on less powerful devices.The second perspective is enhancing predictionaccuracy.Notably,SAM itself lacks specific semantic information,which leads to suboptimal segmentation results in com⁃plex scenarios.Thus,considerable research focuses on enhancing the prediction accuracy of SAM.Subsequently,thestudy thoroughly reviews and analyzes the current applications of the SAM model in various tasks and data types.Theseapplications are divided into three parts:the first part covers applications in image processing-related tasks,including styletransfer,object detection,object counting,image editing,complex image segmentation,and medical image segmentation.However,applying SAM directly to medical image segmentation may not yield satisfactory results,which suggests the needfor further adjustments in specific scenario tasks.The second part encompasses applications in video-related tasks,includ⁃ing video super-resolution,video object tracking,and audio–visual scene segmentation.The third part explores applica⁃tions in other directions,such as point cloud segmentation,3D reconstruction,controllable image caption generation,anddata annotation.Through the organization of the applications of SAM in the three parts,the study summarizes the advan⁃tages and limitations of applying SAM to various downstream tasks.These analyses can assist researchers in better applyingand improving SAM,which enhances its robustness and generalization capabilities.Finally,the study proposes severalvaluable future research directions for the SAM model.These directions include:1)modularization:although SAM hasalready demonstrated excellent performance in certain tasks,its efficiency and flexibility still need to be improved.Withthe continuous expansion of SAM application domains,many applications have put forward the requirement for SAM to pos⁃sess new knowledge.Therefore,the model is required to have domain adaptation and continuous learning capabilities.Drawing inspiration from large language models,new modular structures can be added to SAM to enhance its domain adap⁃tation and continuous learning capabilities.2)Weakly supervised semantic segmentation:in weakly supervised semanticsegmentation,retraining model classification and generating pseudo-labels are typically necessary,but they involve timeconsuming and intricate steps.Recent studies use SAM as a base model in this domain,which capitalizes on its strong gen⁃eralization for satisfactory results without fine-tuning.However,although SAM can produce relatively clear results in manyexplicit scenarios,SAM has difficulty generating accurate segmentation masks in certain semantically ambiguous scenariosbecause its model does not contain semantic information.We can consider using more diverse weak labels for SAM andincorporating additional post-processing modules to enhance the segmentation accuracy of SAM and improve its perfor⁃mance in weakly supervised semantic segmentation for solving the abovementioned complexity.Exploring the application ofSAM as a foundational model in weakly supervised semantic segmentation,which potentially yields promising results.3)Multimodal fusion for image segmentation:at present,the prompt input of SAM mainly includes four forms:point,tar⁃get box,split mask,and text prompt.However,the continuous expansion of the application areas of SAM has introducednew requirements for cue input forms.The current focus of SAM is on 2D visual tasks,with potential consideration forfuture applications in 3D visual tasks.These applications include considering different input modalities for SAM prompts,introducing time-series prompts to address the limitations of SAM in video processing tasks,and further improving the performance of SAM in various video downstream tasks.4)Efficient fine-tuning of SAM:although SAM has been widely usedin various domains,its performance still falls short compared with other state-of-the-art models in the domain in certain spe⁃cific application scenarios.Studies have shown that its performance is improved by fine-tuning SAM for domain-specificdatasets.However,the fine-tuning process is costly due to the large size of the SAM model.Therefore,performing finetuning efficiently becomes an important issue.Given the substantial parameter count of SAM,incorporating new modulesinto the model,freezing its core during training,and only training the newly added modules significantly reduce the train⁃ing cost.This approach facilitates further research on the application of SAM in various downstream tasks.5)Leveraginggestalt psychology’s holistic cognitive perspective to enhance SAM’s adversarial robustness:the vulnerability of SAM toattacks may be due to overfitting on local cognitions.Introducing holistic cognition can prevent overfitting on local cognitionand resist attacks involving noise.By consolidating and summarizing SAM in this study,SAM can be further developed andapplied to drive the advancement of foundational models in the field of computer vision.