A comparative study of GhostStem and C2f in YOLOv5s under near-constant compute

View at publisher
Authors: Xiaoqiang Luo, Wenjing Peng, Rongchang Wei, Haifeng Qian, Xuemei Qin
Cited by: 0

Abstract

In this study, we explore lightweight adjustments to YOLOv5s with near-constant complexity by replacing the stem with GhostStem and C3 with C2f. On Pascal VOC 2007, C2f improved accuracy through better gradient flow and feature reuse, while GhostStem reduced early computation. Yet, their combination did not consistently outperform single modules, suggesting feature conflicts. Complexity remained similar. These findings reveal both the promise and limits of simple module fusion. We will explore alternative placements, integrate attention, expand to VOC2012/COCO, and assess edge hardware deployment.

Fields of Study & Subjects

Advanced Neural Network Applications Domain Adaptation and Few-Shot Learning Visual Attention and Saliency Detection

Publication details

Pages
27–27
Publication date
27 January 2026