Feature Synthesis for Few-Shot Object Detection

The Few-Shot Object Detection (FSOD) task aims to detect novel instances in scenarios with limited data. However, the feature distribution of the novel class can be easily influenced by the distribution of features from the base classes. This paper introduces the Feature Synthesis for Few-Shot Object Detection algorithm, leveraging Generative Adversarial Networks to generate visual features for novel classes. By combining semantic embeddings with real visual features, the generator is trained to enhance the correlation between synthetic features and their corresponding categories. Class prototypes are computed based on real features, and contrastive loss guides the constraint of the synthetic feature distribution, improving model performance. Additionally, the algorithm incorporates Pseudo Margin Evaluation loss to calculate instance uncertainty scores and increase discrimination power. Experimental results on the MS-COCO dataset demonstrate the algorithm’s effectiveness with significant performance gains.