Meta Faster R-CNN: Towards Accurate Few-Shot Object Detection with Attentive Feature Alignment

• DOI: 10.1609/aaai.v36i1.19959
• Publication Date: 2022-07-04T09:04:51Z
• Authors: Guangxing Han; Shiyuan Huang; Jiawei Ma; Yicheng He; Shih-Fu Chang
• Keywords: Discriminative model; Single shot; Feature (linguistics)

Few-shot object detection (FSOD) aims to detect objects using only a few examples. How adapt state-of-the-art detectors the few-shot domain remains challenging. Object proposal is key ingredient in modern detectors. However, quality of proposals generated for classes existing methods far worse than that many-shot classes, e.g., missing boxes due misclassification or inaccurate spatial locations with respect true objects. To address noisy problem, we propose novel meta-learning based FSOD model by jointly optimizing generation and fine-grained classification. improve learn lightweight metric-learning prototype matching network, instead conventional simple linear object/nonobject classifier, used RPN. Our non-linear classifier feature fusion network could discriminative recall classes. classification, attentive alignment method misalignment between thus improving performance detection. Meanwhile separate Faster R-CNN head base show strong maintaining base-classes knowledge. achieves on multiple benchmarks over most shots metrics.