Background Shift

Background shift in continual semantic segmentation (CSS) refers to the ambiguity of background pixels, which can represent either previously unseen or already-learned classes, hindering accurate model updates. Current research focuses on mitigating this issue through techniques like pseudo-labeling refinement, memory-based approaches (e.g., exemplar memory), and improved classifier initialization strategies that leverage background information selectively or transfer relevant weights from previous classes. Addressing background shift is crucial for robust CSS, enabling efficient and accurate learning of new classes in applications like autonomous driving and robotics where continuous model adaptation is essential.