SSP 自支持匹配的小样本分割任务新思想 Self-Support Few-Shot Semantic Segmentation-范琦.pdf

Self-Support Few-Shot Semantic Segmentation Qi Fan HKUST Few-shot Semantic Segmentation Problem: Intra-class appearance variations 2 Qi Fan HKUST CSE Problem • Appearance gap between query and support • Data scarcity: there are very few support images of novel classes • Data diversity: there are infinite query images of novel classes • How to solve it? • Mine the support images? • Mine obj ects of novel classes from the train set? • These methods can not fundamentally solve the problem • Our novel idea: Use query to self-support itself 3 Qi Fan HKUST CSE Related Works • Mining Latent Classes for Few-shot Segmentation • Few-Shot Semantic Segmentation with Cyclic Memory Network • Learning Meta-class Memory for Few-Shot Semantic Segmentation • Adaptive Prototype Learning and Allocation for Few-Shot Segmentation • AMP: Adaptive Masked Proxies for Few-Shot Segmentation • Prototype Mixture Models for Few-shot Semantic Segmentation • Part-aware Prototype Network for Few-shot Semantic Segmentation 4 Qi Fan HKUST CSE Problem • Appearance gap between query and support • Onl