Continual Object Learning Using Gaussian Modeling

Ignacio Romero, Patricio Loncomilla, Javier Ruiz-del-Solar

Abstract: The continual learning of objects is a significant challenge that can be addressed by using separate learning models for feature extraction and object classification, trained at different times. In this paper, we propose implementing this idea using pre-trained foundation models for feature extraction and Gaussian representations for object classification, where Gaussian distributions are used to characterize each object class independently of other classes. Gaussian distributions offer two key advantages in characterizing object classes. First, each object class model operates independently, facilitating the straightforward incorporation of new classes. Second, Gaussians provide a compact representation of object classes, requiring few samples to construct a model that can be refined as new samples become available. The proposed approach is validated in an image classification task using the ImageNet and Caltech 256 datasets.

External IDs:doi:10.1007/978-3-032-11381-8_16