Few-shot object detection via data augmentation and distribution calibration

General object detection has been widely developed and studied over the past few years, while few-shot object detection is still in the exploratory stage. Learning effective knowledge from a limited number of samples is challenging, as the trained model is prone to over-fitting due to biased feature distributions in a few training samples. There exist two significant challenges in traditional few-shot object detection methods: (1) The scarcity of extreme samples aggravates the proposal distribution bias, hindering the evolution of regions of interest (ROI) heads toward new categories; (2) Due to the scarce of the samples in novel categories, the region proposal network (RPN) is identified as a key source of classification errors, resulting in a significant decrease in detection performance on novel categories. To overcome these challenges, an effective knowledge transfer method based on distributed calibration and data augmentation is proposed. Firstly, the biased novel category distributions are calibrated with the basic category distributions; secondly, a drift compensation strategy is utilized to reduce the negative impact on new categories classifications during the fine-tuning process; thirdly, synthetic features are obtained from calibrated distributions of novel categories and added to the subsequent training process. Furthermore, the domain-aware data augmentation is utilized to alleviate the issue of data scarcity by exploiting the cross-image foreground—background mixture to increase the diversity and rationality of augmented data. Experimental results demonstrate the effectiveness and applicability of the proposed method.