A visual inductive priors framework for data-efficient image classification.

Pengfei Sun, Xuan Jin, Wei Su, Yuan He, Hui Xue', Quan Lu

Keywords: learn from scratch, classification, visual inductive priors, DSK-net, induced hierarchy

Abstract: State-of-the-art classifiers rely heavily on large-scale datasets, such as ImageNet, JFT-300M, MSCOCO, Open Images, etc. Besides, the performance may decrease significantly because of insufficient learning on a handful of samples. We present Visual Inductive Priors Framework (VIPF), a framework that can learn classifiers from scratch. VIPF can maximize the effectiveness of limited data. In VIPF, we propose a novel image classification architecture, called Dual Selective Kernel network(DSK-net), which is robust to translation invariance. With more discriminative feature extracted from DSK-net, overfitting of network is lightened. Then a loss function based on positive class is applied for model training. Additionally, an induced hierarchy is used which is easier for VIPF to learn from scratch. Finally, we won the 1st place in VIPriors image classification competition.