One-Shot Object Detection with 4D-Correlation and 4D-Attention

One-shot object detection (OSOD) is an emerging and important task to detect novel classes given only one query sample. This task is challenged by the significant variability in appearance and geometry within query-target image pairs, which blurs the lines between variations of the inter- and intra classes. To address such issue, we propose a novel framework called 4D-Correlation and 4D-Attention Networks (4DCA) to enhance the differentiation of these variations. Our 4DCA framework consists of two collaborative components: (1) During feature extraction, the 4D-Correlation-Pyramid (FCP) module dynamically fuses multi-layer features to construct high-dimensional representations that accentuate the correlations between query and target images. (2) For correlation learning, the 4D-Attention (FA) module employs multi-head cross-attention mechanisms learn the complex correlation between query-target image pairs. Our method focuses on extracting abundant visual correlations in limited samples, which strengthens intra-class robustness and inter-class separability. Extensive experiments on the MS COCO and FSOD datasets have proven the superiority of our method.