Few-Shot Object Detection Algorithm Based on Adaptive Relation Distillation

Dan-Ting Duan, Wei Zhong, Liang Peng, Shuang Ran, Fei Hu

Abstract: Deep learning methods have advanced the accuracy and speed of object detection. However, acquiring labeled data is especially challenging in real-world scenarios. As a result, the general-purpose object detection algorithms based on deep learning experience significant performance degradation with limited samples. To tackle the issue of declining detection accuracy in the presence of few-shot data, this paper introduces a few-shot object detection algorithm based on adaptive relation distillation, which improves upon existing algorithms by enhancing the fusion of query and support features. In the proposed method, the adaptive relational distillation module discards the hand-designed and inefficient query information utilization strategy employed in previous algorithms and adaptively fuses the features of support and query images using convolutional networks. To augment the learning capability of the adaptive relational distillation module, we utilize a hybrid attention module in the support branch to emphasize the regions crucial for detecting specific classes of objects. The experimental results demonstrate our proposed algorithm achieves an average accuracy of 47.6% on three data divisions and five sample size settings for the PASCAL VOC dataset, which marks an improvement of 8.3% over DCNet.

External IDs:dblp:conf/prcv/DuanZPRH23