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.
© 2026 Elsevier Inc. Few-shot Object Counting (FSC) aims at estimating the number of instances from novel categories given only a few annotated exemplars. Although existing approaches have achieved notable progress, most of them treat exemplars as holistic entities, making them vulnerable to occlusion, deformation, and appearance ambiguity in real-world scenes. To address these challenges, we decide to Dive into the Details (DDCounter) and propose the holistic and partial feature fusion network, a novel FSC framework that introduces a holistic–partial exemplar fusion strategy and feature alignment for robust density estimation. Specifically, the Prior Mask Generation module based on the Segment Anything Model provides precise object-aware masks to suppress background noise. A Scale Embedding module further injects spatial-aware scale cues to enhance object size sensitivity. We then adopt a Holistic–Partial Feature Fusion strategy, where exemplar features are decomposed into discriminative local tokens via random sampling and complemented with holistic global features for robust target representation. These enriched exemplar features are used to guide query feature enhancement through an attention-based Feature Alignment module. A Density Regression Head finally outputs high-resolution density maps for accurate counting. Extensive experiments on FSC-147 and CARPK demonstrate that DDCounter consistently outperforms state-of-the-art methods, particularly in high-density or visually complex scenarios.
Author(s): Wu W, Wang S, Zhang H
Publication type: Article
Publication status: Published
Journal: Journal of Visual Communication and Image Representation
Year: 2026
Volume: 117
Print publication date: 01/04/2026
Online publication date: 09/03/2026
Acceptance date: 07/03/2026
ISSN (print): 1047-3203
ISSN (electronic): 1095-9076
Publisher: Academic Press Inc.
URL: https://doi.org/10.1016/j.jvcir.2026.104773
DOI: 10.1016/j.jvcir.2026.104773
Altmetrics provided by Altmetric
