TY  - CONF
AU  - JW Xiao
AU  - CB Zhang
AU  - J. Feng
AU  - Xialei Liu
AU  - Joost Van de Weijer
AU  - MM Cheng
A2  - CVPR
PY  - 2023//
TI  - Endpoints Weight Fusion for Class Incremental Semantic Segmentation
BT  - Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition
SP  - 7204
EP  - 7213
N2  - Class incremental semantic segmentation (CISS) focuses on alleviating catastrophic forgetting to improve discrimination. Previous work mainly exploit regularization (e.g., knowledge distillation) to maintain previous knowledge in the current model. However, distillation alone often yields limited gain to the model since only the representations of old and new models are restricted to be consistent. In this paper, we propose a simple yet effective method to obtain a model with strong memory of old knowledge, named Endpoints Weight Fusion (EWF). In our method, the model containing old knowledge is fused with the model retaining new knowledge in a dynamic fusion manner, strengthening the memory of old classes in ever-changing distributions. In addition, we analyze the relation between our fusion strategy and a popular moving average technique EMA, which reveals why our method is more suitable for class-incremental learning. To facilitate parameter fusion with closer distance in the parameter space, we use distillation to enhance the optimization process. Furthermore, we conduct experiments on two widely used datasets, achieving the state-of-the-art performance.
UR  - http://dx.doi.org/10.1109/CVPR52729.2023.00696
N1  - LAMP
ID  - JW Xiao2023
ER  -