Research Status and Analysis of Few-Shot Semantic Segmentation
The traditional semantic segmentation task is usually data-driven,which requires large-scale intensive annotation data for training.And it cannot be generalized to novel class segmentation which means that when novel class samples emerge,it is necessary to collect a large amount of data to retrain the model.So these issues severely limit its practical application.In order to alleviate the issues of data scarcity and poor generalization ability in traditional semantic segmentation task,the few-shot semantic segmentation(FSS)task has been proposed,which can utilize the past accumulated knowledge to achieve the segmentation of novel classes in images that are not included in the training process with only a limited number of densely labeled samples.FSS can significantly reduce the need for data annotation and demonstrates good generalization performance.It has attracted widespread attention from the academic community.This research filed is crucial for practical applications where data availability is limited and difficult,such as medical field,remote sensing filed and so on.In recent years,there has been rapid development and progress in the field of few-shot semantic segmentation.A large number of novel and high-performing methods have emerged,urgently requiring a comprehensive review to summarize and sort out these algorithms.This paper investigates the research on few-shot semantic segmentation in the field of natural images from the aspects of basic knowledge,algorithms summarization,extended applications and future development.We firstly introduce the basic knowledge of few-shot semantic segmentation task,including its origin,core ideas,conceptual knowledge,challenges,dataset benchmarks,and evaluation metrics.Then we have analyzed,compared the current few-shot semantic segmentation methods in detail,and summarized them accordingly.Specifically,we divide those methods into optimization-based and metric-learning-based methods according to whether there is gradient backpropagation during inference.When it comes to optimization-based methods,there is a need for gradient backpropagation to optimize the model during inference,whereas metric-learning-based approaches do not require any optimization of the model during inference.Instead,they utilize densely labeled images to obtain information about the categories to be segmented and use this class-specific information to guide the segmentation process.This allows for a more efficient and direct utilization of the labeled data without the need for further model training or optimization.We have conducted a comparison of the introduced few-shot semantic segmentation algorithms,listing the best performance achieved by each algorithm and providing quantitative analysis on three datasets.The comparison was made from three perspectives:the performance comparison on each dataset under different evaluation metrics,the performance comparison of a single algorithm across different dataset benchmarks,and the results comparison of different algorithms on the PASCAL-5i dataset.In addition,we introduce some tasks of integrating few-shot semantic segmentation and other technologies,such as few-shot instance segmentation,generalized few-shot semantic segmentation,incremental few-shot semantic segmentation,weakly supervised few-shot semantic segmentation,and cross-domain few-shot semantic segmentation.Finally,we discuss the existing problems and future development trend of few-shot semantic segmentation.We believe that this paper will help researchers better understand the current research status of few-shot semantic segmentation,provide a broader perspective for future studies,and promote further development and innovation in this field.
NETL
NSTL
万方数据
