Enhancing few-shot object detection through mixing and separating tuning strategies

Abstract: Few-shot object detection (FSOD) aims to detect novel objects using only a limited number of labelled examples. Existing fine-tuning-based FSOD methods typically face challenges in effectively transferring knowledge from base to novel categories, often leading to confusion between them. To address this issue, we propose a novel mixing and separating tuning (MST) framework. In the mixing tuning stage, we pretune the model using transitional samples between base and novel categories to reduce bias towards the base category. Subsequently, in the separating tuning stage, we further fine-tune the model on novel category samples with an auxiliary discrimination network and an energy-based separation strategy. Extensive experimental results on PASCAL VOC and Microsoft COCO benchmarks demonstrate that our MST framework significantly outperforms existing state-of-the-art methods, achieving better discrimination and separation between base and novel categories. The proposed approach not only improves detection performance on novel categories but also maintains high accuracy on base categories. The code is available at: http://github.com/zqpiao/MS_FSOD.