Image Retrieval Under Fine-Grained and Long-Tailed Distribution

Shuai Chen, Fanman Meng, Qingbo Wu, Yuxuan Liu, yaofeng yang

Abstract: Image retrieval is a promising task that aims to retrieve relevant images from a large-scale database based on user’s requests. This paper is a solution to the Huawei 2020 digital device image retrieval competition. The core idea of the proposed solution is to employ metriclearning to perform fine-grained image retrieval. To be specific, to address the long-tailed distribution caused by the imbalance of samples in the dataset, an image retrieval tailored causal graph is first constructed, and a causal intervention is performed for counterfactual reasoning, which proves to be effective to alleviate the influence of long-tailed distribution. To solve the challenging fine-grained image retrieval issues, this paper proposes a novel global and local attention image retrieval framework, which simultaneously mines global and local features to obtain the most discriminative feature. In addition, an object detector is further developed to capture the object of interest, thereby an accurate representation of the foreground area can be acquired. Furthermore, some additional testing and model ensemble skills, such as re-ranking, finetuning on larger images, and multi-scale testing, are implemented to further boost the performance. Extensive experiments on the benchmark demonstrate the effectiveness of the proposed method.