Few-Shot Object Detection via Back Propagation and Dynamic Learning

Abstract

Utilizing traditional object detectors to build a few-shot object detection (FSOD) model ignores the differences between classification and regression tasks and causes task conflict and class confusion, resulting in a decline in classification performance. In contrast, this paper focuses on the above shortcomings and utilizes the strategies of Back Propagation and Dynamic Learning to construct a model for addressing FSOD, named BPDL. Our BPDL has a two-fold main idea: a) it uses the optimized localization boxes to alleviate the task conflict and refine classification features by a correction loss, and b) it develops a dynamic learning strategy to filter the confusing features and mine more realistic prototype representations of the categories to calibrate classification. Extensive experiments on multiple benchmarks show that our BPDL model outperforms existing methods and advances the FSOD task’s state-of-the-art.