Generalized Feature Learning for Detection of Novel Objects

Few-shot object detection (FSOD) aims at heuristically detecting novel objects with limited labeled data. Typical methods focus on the advanced classifications using the features extracted from common backbones. However, these features are usually base domain-biased, which trap the methods due to insufficient knowledge learned by common backbones. In this correspondence, a novel FSOD network is designed via learning of deep-level generalized features. Specifically, a two-branch backbone is introduced by adding a category-agnostic feature extractor as a parallel branch of common backbone, which preserves valuable but coarse features out of base classes. To fully refine these features, a source aggregation scheme with probabilistic pathway selection and source-based channel dropout is designed, which prevents the network from falling into the dominant optimization of base classes. The resulting generalized features are less-biased features, which increases the adaptability to novel classes. Besides, a loose contrastive loss is provided as extra supervised information to relieve overfitting. As a result, the proposed method reaches a good compatibility with the data out of base classes. The effectiveness of the proposed method is verified through experiments on PASCAL VOC, COCO, and iCubWorld Transformations datasets.