Congratulations to Hao Li and Yuhang Xia on Their AAAI 2025 (CCF-A) Paper Acceptance.

AAAI 2025 (CCF-A, Top Conference in Artificial Intelligence)

Paper Title: Deconfound Semantic Shift and Incompleteness in Incremental Few-shot Semantic Segmentation

Paper Code:https://github.com/LegendSherlock/Deconfound-IFSS Abstract: Incremental Few-Shot Semantic Segmentation (IFSS) extends the segmentation capability of trained models, enabling them to segment images of new classes with few samples. However, during the incremental learning process, semantics may shift between background and object classes or vice versa. Additionally, when new classes differ significantly from pre-learned old classes, samples from new classes often lack representative feature attributes. In this paper, we propose a causal framework to discuss the causes of semantic shifts and incompleteness in IFSS and eliminate the revealed causal effects from two aspects. First, we introduce a Causal Intervention Module (CIM) to resist semantic shifts. CIM gradually and adaptively updates the prototypes of old classes, intervening to remove confounding factors. Secondly, we propose a Prototype Refinement Module (PRM) to complete missing semantics. In the PRM, knowledge obtained from the scene learning scheme helps integrate the features of new and old class prototypes. Experiments on the PASCAL-VOC 2012 and ADE20k benchmarks demonstrate the superior performance of our method.