Browse by author
Lookup NU author(s): Dr Shidong WangORCiD
Full text for this publication is not currently held within this repository. Alternative links are provided below where available.
© 2025 Elsevier B.V.Few-shot Object Counting (FSC) aims to accurately count objects of arbitrary categories in the query images. The standard pipeline is to extract exemplar features from the feature of query image and match them to obtain the final object counts. However, query image and exemplars often contain excessive background information and biases that do not belong to a specific category, which compromises the performance of feature matching and object counting. Another problem for traditional methods is that the ground truth density map is generated with fixed-size Gaussian distribution, which is inconsistent with the objects of varying scales in actual images. To address these problems, we propose a framework, termed as Cross REfinement and Adaptive density Map (CREAM), to extract the foreground information of all exemplars, eliminate the background information and make the single exemplar feature unbiased for the category to be counted. And in the same way the exemplars are used to weed out the background information in the query image. Moreover, instead of uniformly labeling all objects with Gaussian density maps of the same scale, we design a novel algorithm to generate the ground truth density map that considers the correlation between the density distribution of a single object and its scale, which is more consistent with the realistic scenarios. Extensive experiments on large-scale datasets, such as FSC-147, CARPK and ShanghaiTech, show that our method significantly outperforms the state-of-the-art approaches, and the detailed analysis also shows the effectiveness of each designed module. Code is available at https://github.com/CBalance/CREAM.
Author(s): Xu Y, Li M, Ye Q, Wang S, Li L, Zhang H
Publication type: Article
Publication status: Published
Journal: Image and Vision Computing
Year: 2025
Volume: 161
Print publication date: 01/09/2025
Online publication date: 29/06/2025
Acceptance date: 17/06/2025
ISSN (print): 0262-8856
ISSN (electronic): 1872-8138
Publisher: Elsevier Ltd
URL: https://doi.org/10.1016/j.imavis.2025.105632
DOI: 10.1016/j.imavis.2025.105632
Altmetrics provided by Altmetric
