DR-CIML: Few-shot Object Detection via Base Data Resampling and Cross-iteration Metric Learning

Abstract

Aiming at the problems of difficult data collection and labor-intensive manual annotation, few-shot object detection (FSOD) has gained wide attention. Although current transfer-learning-based detection methods outperform meta-learning-based methods, their data organization fails to fully utilize the diversity of the source domain data. In view of this, Data Resampling (DR) organization is proposed to fine-tune the network, which can be employed as a component of any model and dataset without additional inference overhead. In addition, in order to improve the generalization of the model, a Cross-Iteration Metric-Learning (CIML) branch is embedded in the few-shot object detector, thus actively improving intra-category feature propinquity and inter-category feature discrimination. Our generic DR-CIML approach obtained competitive scores in extensive comparative experiments. The nAP50 performance on PASCAL VOC improved by up to 6.3 points, and the bAP50 performance reached 81.0, surpassing the base stage model (80.8) for the first time. The nAP75 performance on MS COCO improved by up to 1.6 points.