Variational Multi-Scale Feature Integration for Few- Shot Object Detection

Traditional few-shot object detection(FSOD) models are typically trained on benchmark datasets such as PASCAL VOC 2007+2012 and MS COCO 2014. However, these datasets exhibit significant limitations in terms of scale distribution and target density, making it challenging for models trained on them to perform well in multi-scale scenarios, especially when detecting dense small objects. As a result, models trained on these datasets often struggle with sensitivity to scale variations, reducing their effectiveness in handling complex multi-scale object detection tasks. To address this issue, we utilize the VisDrone2019 dataset, which features a wide range of scale distributions and significant scale variability, and construct a FSOD dataset from it. Based on this dataset, we integrate Variational Autoencoders (VAE) and their variants into the classical FSOD framework Meta R-CNN. This integration significantly enhances the model's capability to model fine-grained feature distributions of support samples and improves robustness through variational feature generation. Furthermore, we incorporate Feature Pyramid Networks (FPN) and their variants to optimize multi-scale object feature representations, achieving notable improvements in performance for multi-scale object detection. Experimental results demonstrate that our method outperforms existing state-of-the-art (SOTA) approaches on the VisDrone2019 few-shot object detection dataset, showcasing significant advantages in addressing multi-scale scenarios.