Few-shot object detection with dense-global feature interaction and dual-contrastive learning

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 53; no. 11; pp. 14547 - 14564
• Main Authors: Huang, Lian, Dai, Shaosheng, He, Ziqiang
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.06.2023; Springer Nature B.V
• Subjects: Artificial Intelligence; Computer Science; Embedding; Learning; Localization; Machines; Manufacturing; Mechanical Engineering; Methods; Modules; Neural networks; Object recognition; Processes; Semantics; Source code; Few-shot object detection; Supervised contrastive learning; Convolutional neural networks; Attention
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-022-04243-3

Few-shot object detection (FSOD) aims to detect novel objects quickly from extremely few annotated examples of previously unseen classes. Most existing methods follow the meta-learning paradigm and still encounter two typical challenges: how to interact information efficiently and how to learn a good decision boundary in the embedding space. To address these challenges, we propose an FSOD method with dense-global feature interaction and dual-contrastive learning. Specifically, the well-designed dense-global feature interaction (DGFI) module integrates dense spatial attention and global context attention to capture the correlation across the support images and query images. The DGFI module then leverages the correlation to yield the interactive features. The dual-contrastive learning (DaCL) module, which consists of supervised contrastive branches and prototypical contrastive branches, incorporates object-level contrastive learning to endow learned features with good intra-class similarity and inter-class distinction. The more discriminative features form a well-separated decision boundary in the embedding space and alleviate the typical misclassification problem. Extensive experiments on the PASCAL VOC, MS-COCO and Object365 benchmarks demonstrate that the proposed method achieves state-of-the-art performance. Source code is publicly available at https://github.com/HuangLian126/DGFIDaCL .