Instance Feature Caching for Cross-Domain Few-Shot Object Detection

Yali Huang, Jie Mei, Yiming Yang, Mi Guo, Mingyuan Jiu, Mingliang Xu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 1567-1575

Deep learning-based object detection has made significant progress owing to large-scale labeled data. However, sufficient labeled data is unavailable in many practical scenarios, leading to poor model performance. Cross-Domain Few-Shot Object Detection aims to solve the above problem by utilizing the knowledge (i.e. pre-trained model or data) of the source domain with large labeled data to improve the detection performance in the target domain with few data. This task is very challenging, because data distributions of the source and target domains may be significantly different, failing to achieve performance generalization. In this paper, we propose a novel cross-domain detection method with Instance-Feature-Caching. Firstly, we utilize a few labeled data in the target domain and their corresponding tag encodings to build an instance caching with a set of rich semantic features. Secondly, the adaptive adjustment of the encoding in the detector is achieved through the instance retrieval mechanism. Finally, in order to boost the adaptability and generalization ability of the instance caching, we design an instance update strategy to dynamically learn the instance features in the caching. Experimental results show that the proposed method outperforms existing methods on several standard cross-domain few-shot object detection benchmarks, verifying its effectiveness and robustness in the cross-domain and few-shot object detection task.