MINI: Mining Implicit Novel Instances for Few-Shot Object Detection

• DOI: 10.48550/arxiv.2205.03381
• Publication Date: 2022-01-01
• Authors: Yuhang Cao; Jiaqi Wang; Yiqi Lin; Dahua Lin
• Keywords: Discriminative model; Pascal (unit); Base (topology); Labeled data; Transfer of learning

Learning from a few training samples is desirable ability of an object detector, inspiring the explorations Few-Shot Object Detection (FSOD). Most existing approaches employ pretrain-transfer paradigm. The model first pre-trained on base classes with abundant data and then transferred to novel annotated samples. Despite substantial progress, FSOD performance still far behind satisfactory. During pre-training, due co-occurrence between classes, learned treat co-occurred as backgrounds. transferring, given scarce suffers learning discriminative features distinguish instances backgrounds classes. To overcome obstacles, we propose framework, Mining Implicit Novel Instances (MINI), mine implicit auxiliary samples, which widely exist in but are not annotated. MINI comprises offline mining mechanism online mechanism. leverages self-supervised collaboratively trained network. Taking mined takes teacher-student framework simultaneously update network fly. Extensive experiments PASCAL VOC MS-COCO datasets show achieves new state-of-the-art any shot split. significant improvements demonstrate superiority our method.