Given a query patch from a novel class, one-shot object detection aims to detect all instances of that class in a target image through the semantic similarity comparison. However, due to the extremely limited guidance in the novel class as well as the unseen appearance difference between query and target instances, it is difficult to appropriately exploit their semantic similarity and generalize well. To mitigate this problem, we present a universal Cross-Attention Transformer (CAT) module for accurate and efficient semantic similarity comparison in one-shot object detection. The proposed CAT utilizes transformer mechanism to comprehensively capture bi-directional correspondence between any paired pixels from the query and the target image, which empowers us to sufficiently exploit their semantic characteristics for accurate similarity comparison. In addition, the proposed CAT enables feature dimensionality compression for inference speedup without performance loss. Extensive experiments on COCO, VOC, and FSOD under one-shot settings demonstrate the effectiveness and efficiency of our method, e.g., it surpasses CoAE, a major baseline in this task by 1.0% in AP on COCO and runs nearly 2.5 times faster. Code will be available in the future.
@article{arxiv.2104.14984,
title = {CAT: Cross-Attention Transformer for One-Shot Object Detection},
author = {Weidong Lin and Yuyan Deng and Yang Gao and Ning Wang and Jinghao Zhou and Lingqiao Liu and Lei Zhang and Peng Wang},
journal= {arXiv preprint arXiv:2104.14984},
year = {2021}
}