Critic Boosting Attention Network on Local Descriptor for Few-shot Learning

Chengzhang Shi, Chung-Ming Own, Ching-Chih Chou, Bailu Guo

Abstract: Few-shot learning is a new upsurge of research in the computer vision field to adapt to new tasks with only a few examples. Recent works commonly use meta-learning or metric learning methods to solve the problem. In contrast, many of them usually use image-level feature representations and treat all parts of the image equally. This manipulation cannot cast off the cluttered background's interference or grasp the key-points of the idea rapidly. In our study, the authors use local descriptors assisted with image-to-class metric learning measures and utilize a critic following the human visual system mechanism to force the model to concentrate on the image's essential semantic parts. The unsupervised critic gives each of the local descriptors of an input image a score evaluating its significance and contribution to the final classification. Thus the model can relieve from the image background cluttering and fast focus on the essential parts. Generally, our proposed critic boosting attention network (short as CBAnet) can be classified as a kind of attention mechanism. Still, we do not need to use the background impact in an unsupervised way. The extensive experiments that we conduct on several benchmarks testify to our proposed method's superiority for few-shot learning.