CineFault-DG: Domain-Generalized Multimodal Fault Diagnosis for Quality Control in Virtual Film Production
IEEE Access · 2026 · 14 · 115112-115125
DOI: 10.1109/access.2026.3710241 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yining Xu und 3 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Jiangsu Provincial Graduate Research and Practice Innovation Program.
Abstract (Originalsprache)
Virtual production has become a central workflow in contemporary film and television making, combining LED volume stages, real-time rendering, camera tracking, color management, and post-production tools into a tightly coupled creative-technical pipeline. However, the same coupling that enables in-camera visual effects also creates a new class of production quality failures, including LED refresh banding, tracking-render parallax, color-space mismatch, genlock jitter, rolling-shutter artifacts, focus and exposure drift, and compression-induced temporal instability. Existing video quality assessment methods estimate perceptual quality, while intelligent cinematography research mainly focuses on shot generation, editing, camera control, or semantic understanding. Few studies formulate virtual production quality control as a fault diagnosis problem with cross-scene generalization and open-set anomaly rejection. This paper proposes CineFault-DG, a domain-generalized multimodal fault diagnosis framework for virtual film production quality control. The method combines a spatiotemporal visual encoder, production-aware quality descriptors, scene-style disentanglement, graph-guided fault reasoning, and uncertainty-calibrated open-set detection. To support reproducible evaluation, we define a construction protocol for CineVP-QA, a benchmark derived from public video quality and movie-understanding datasets with physically motivated virtual-production fault injection. As an initial proof of concept, we implement a public-data pilot using 1,920 synthetic short clips generated from eight open sample-image domains. In this pilot, a lightweight diagnostic implementation reaches 0.981 random-split macro-F1, while performance drops to 0.160 macro-F1 under a held-out-domain split, revealing the importance of cross-domain evaluation. We further evaluate a real YouTube-UGC H264 subset under a 200 MB download budget. This subset contains 14 videos, 420 generated clips, two injected severity levels, and seven content categories; it yields 0.849 accuracy, 0.825 macro-F1, 0.116 severity mean absolute error, and 0.706 open-set AUROC. Robustness analyses further report ten-run mean and standard deviation estimates, metadata-context behavior, loss-weight sensitivity, and open-set threshold behavior. CineFault-DG provides a practical bridge between film production studies and intelligent maintenance research by treating creative media workflows as diagnosable production systems.
Förderung
- National Natural Science Foundation of China
- Jiangsu Provincial Graduate Research and Practice Innovation Program
Zitieren
APA
Xu, Y., Wang, F., Cao, Y., & Wang, Y. (2026). CineFault-DG: Domain-Generalized Multimodal Fault Diagnosis for Quality Control in Virtual Film Production. IEEE Access, 14, 115112-115125. https://doi.org/10.1109/access.2026.3710241
Vancouver
Xu Y, Wang F, Cao Y, Wang Y. CineFault-DG: Domain-Generalized Multimodal Fault Diagnosis for Quality Control in Virtual Film Production. IEEE Access. 2026;14:115112-115125. doi:10.1109/access.2026.3710241
BibTeX
@article{Xu2026,
author = {Xu, Yining and Wang, Fang and Cao, Yajie and Wang, Yongjian},
title = {CineFault-DG: Domain-Generalized Multimodal Fault Diagnosis for Quality Control in Virtual Film Production},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {115112--115125},
doi = {10.1109/access.2026.3710241}
}
EUVIMED