Faster Few-Shot Object-Detection: Two-Stage Fine Tuning Integrated with External Memories
Yunpu Zhang, Zhuowei Wang, Shichao Zhou, Yingrui Zhao · 2024
Traditional object detection requires a large amount of annotated data to ensure model accuracy and generalization, but in practical scenarios, labeled samples are often limited. Few-shot Object Detection (FSOD) addresses this issue by enabling models to quickly learn and infer with minimal labeled data, which is crucial for real-time applications. Although various FSOD approaches, including feature representation, meta-learning, Generative Adversarial Networks (GANs), and self-supervised learning, have been proposed, challenges like overfitting and poor generalization persist. This paper introduces a two-stage fine-tuning model incorporating an External Memory Module. Through the use of two small, learnable, shared memory units that extract visual word representations from limited images, the model enhances generalization and reduces computational complexity. The external attention mechanism, optimized through back-propagation, improves the model’s effectiveness and efficiency.