Enhanced Class Incremental Semantic Segmentation with Initial Model Pre-tuning

Class Incremental Semantic Segmentation (CISS) enables models to learn new classes sequentially while preserving knowledge of previously learned classes, addressing real-world scenarios with data streams. In this paper, we use a knowledge distillation strategy to solve this challenge, leveraging the knowledge of a model learned from old data. Our method utilizes selective pseudolabeling paired with adaptive feature distillation, focusing exclusively on capturing trustworthy knowledge from past iterations. We propose an approach that uses the old model to initialize the new model, replacing background weight transfer with a transformation-based initialization that leverages all old classifiers. By learning importance and projection matrices during a pre-tuning stage, our method generates distinct new class weights aligned with the feature extractor, mitigating background shift and reducing training complexity. The experimental results on the Pascal VOC 2012 dataset in mean Intersection over Union (mIoU) demonstrate the effectiveness of our proposed method.