Abstract:Few-shot segmentation is an emerging and intriguing subfield within computer vision that tackles the challenging task of segmenting objects or regions in images when only a very limited number of annotated examples are available for each category. Although some existing few-shot segmentation techniques have made notable strides in performance, the task of few-shot segmentation continues to pose significant challenges. Particularly in the context of delving into the actual pixel relation between support and query images, the relation often suffers from the interference due to complicated appearance, shape difference, etc. However, the accuracy of pixel relation concerns the effectiveness of category information transmission, which is crucial for segmenting the target objects. In addition, most models are prone to overfitting observed base classes during the training process, resulting in the learned model being unable to be generalized to a wider range of invisible classes. To this end, a novel framework named Combining Hierarchical Sparse Representation with Adaptive Prompt for Few-shot Segmentation (HSRap) is proposed to mine more precise pixel relations between the support and query images, enhancing the transfer of category information and improving the generalization ability of the model. Specifically, HSRap begins with the design of a hierarchical sparse representation module, which uncovers latent dense relation representations between the support and query images through the reconstruction of dense relations with sparse regularization. Following this, a lightweight adaptive prompt meta-learner is designed to generate multiple prompts for each category, ensuring generalizability to a broader range of unseen classes within the same dataset and reducing sensitivity to class shift. HSRap undergoes extensive testing on three challenging benchmarks: PASCAL-5i, COCO-20i, and FSS-1000. The results demonstrate a significant enhancement in the performance of the baseline, achieving competitive results compared to state-of-the-art methods.

Learnable Prompt for Few-Shot Semantic Segmentation in Remote Sensing Domain

Generalized Few-Shot Semantic Segmentation in Remote Sensing: Challenge and Benchmark

Iterative Few-shot Semantic Segmentation from Image Label Text

PartSeg: Few-shot Part Segmentation via Part-aware Prompt Learning

Visual Prompting for Generalized Few-shot Segmentation: A Multi-scale Approach

Self-guided Few-shot Semantic Segmentation for Remote Sensing Imagery Based on Large Vision Models

APSeg: Auto-Prompt Network for Cross-Domain Few-Shot Semantic Segmentation

Combining Hierarchical Sparse Representation with Adaptive Prompt for Few-Shot Segmentation

Semantic Prompt for Few-Shot Image Recognition

Sam-Rsp: A New Few-Shot Segmentation Method Based on Segment Anything Model and Rough Segmentation Prompts

Generalized Few-Shot Meets Remote Sensing: Discovering Novel Classes in Land Cover Mapping via Hybrid Semantic Segmentation Framework

Semi-Supervised Contrastive Learning for Few-Shot Segmentation of Remote Sensing Images

Few-shot Class-Incremental Semantic Segmentation via Pseudo-Labeling and Knowledge Distillation

Prompt-and-Transfer: Dynamic Class-aware Enhancement for Few-shot Segmentation

Insight Any Instance: Promptable Instance Segmentation for Remote Sensing Images

Few-Shot Segmentation of Remote Sensing Images Using Deep Metric Learning

RSPrompter: Learning to Prompt for Remote Sensing Instance Segmentation Based on Visual Foundation Model

Adaptive Self-Supporting Prototype Learning for Remote Sensing Few-Shot Semantic Segmentation

Not Just Learning from Others but Relying on Yourself: A New Perspective on Few-Shot Segmentation in Remote Sensing

One-shot and Partially-Supervised Cell Image Segmentation Using Small Visual Prompt