DF-FSOD: A Novel Approach for Few-shot Object Detection via Distinguished Features

Abstract

Few-shot object detection (FSOD) is a challenging task in which detectors are trained to recognize unseen objects with limited training data. The majority of existing methods are evaluated on the benchmarks built with a fixed quantity of base and novel classes categories. To be specific, the number of base classes is larger than the novel ones. This positively affects the performance evaluated on novel data. However, there are not many works focusing on the effect of such dominated categories on the performance of FSOD models. In this paper, we investigate the efficiency of the detectors in different ratios of base and novel categories in the novel phase. Based on our findings of the affection between base and novel classes, we present a new approach: Distinguished Features for FSOD (DF-FSOD), which encourages the detector to learn distinguished features to capture novel objects via base-class expansion better. In the end, our proposed method outperforms average 4% AP@50 on PASCAL VOC compared to the previous works on the unseen classes when extremely scare labeled data.