M-GAID: A Real-World Dataset for Ghosting Artifact Detection and Removal in Mobile Imaging

Muneeb A. Khan, Hyungsub Kim, Jiho Eum, Yihyun Myung, Yujin Choi, Heemin Park; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 1502-1511

Mobile photography faces unique challenges, including noise, over-saturation, and limited dynamic range due to the constraints of small sensors and apertures. While multiframe fusion techniques enhance image clarity, they often introduce ghosting artifacts in dynamic scenes. Existing DSLR-based datasets fail to address mobile-specific issues like higher noise, lower dynamic range, and real-world conditions. To address this, we present the Mobile Ghosting Artifact Imaging Dataset (M-GAID), the first annotated dataset specifically targeting ghosting artifacts in mobile imaging. M-GAID comprises 2,520 images, including 1,060 with high-frequency 1 and 1,460 with low-frequency ghosting artifacts 2, captured across 87 real-world scenarios with diverse lighting and motion conditions using various mobile devices. Our key contributions are: (1) Ghosting Artifact Focus: M-GAID focuses on high and low-frequency ghosting artifacts specific to mobile photography; (2) Comprehensive Annotations: Images are annotated at the 224x224 pixel patch level, categorizing artifacts by type and location, facilitating precise evaluation of detection and mitigation algorithms; (3) Broad Artifact Spectrum: The dataset covers a range of artifacts, from subtle to severe artifacts, providing a robust benchmark for computational imaging assessments. Evaluations of state-of-the-art models using M-GAID highlight significant challenges, emphasizing its potential to drive advancements in computational imaging solutions tailored for mobile devices.