A spatially aware global and local perspective approach for few-shot incremental learning.

The ability to continuously perceive new concepts with extremely limited samples is innate in human beings. Few-shot incremental learning impersonates this ability by constructing an intelligent learning mechanism, and its intention is to identify novel categories from a given few instances gradually. The crucial means of few-shot incremental learning is to yield a model with the generalization ability to highlight the dominant object in the image and make the appearance of the object smoother by leveraging prior knowledge. In light of this, in this paper, we propose a Spatially Aware Global and Local Perspectives (SGLP) approach to tackle the few-shot incremental learning problem. To enhance semantic representations of features, we build the relationship information of the spatial feature in the global scope and encourage the model to pay attention to the dominant region in features. Furthermore, we assume that the current and surrounding information of the image have a similar appearance and design a smooth operation of the spatial feature by adopting the simple Gaussian kernel in a local scope. Extensive experiments on benchmarks demonstrate the superiority and effectiveness of the proposed approach.