An Empirical Analysis of MIL Pooling Under UCF‐Crime‐to‐ShanghaiTech Transfer in Weakly Supervised Video Anomaly Detection
Applied Computational Intelligence and Soft Computing · 2026-01 · 2026 (1)
DOI: 10.1155/acis/1607118 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Computational Intelligence and Soft Computing, online erschienen am 05.11.2026. Verfasst von Marwan Tayseer Al Zawahra und 2 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Weakly supervised video anomaly detection (WSVAD) reduces the need for dense temporal annotations, but the behavior of Multiple Instance Learning (MIL) aggregation mechanisms under dataset shift remains insufficiently characterized. This study presents a controlled empirical analysis, rather than a new detection architecture, of Mean, Top‐k, and Attention‐based Pooling using I3D and R3D pre‐extracted representations in one transfer direction from UCF‐Crime to ShanghaiTech. Two protocols are distinguished: source‐only cross‐dataset evaluation and transductive unsupervised domain adaptation, in which the same unlabeled ShanghaiTech target videos are available during alignment and subsequent evaluation, while their anomaly labels and frame masks are used only for metric computation. Across five seeds, source‐domain performance is reported as UCF‐Crime video‐level ROC‐AUC, whereas target‐domain localization is evaluated using ShanghaiTech frame‐level ROC‐AUC and Average Precision (AP). For source‐only I3D models, Top‐k Pooling obtained the highest UCF‐Crime video‐level AUC (0.8865 ± 0.0006), whereas Mean Pooling obtained the highest ShanghaiTech frame‐level AUC (0.5749 ± 0.0097) and Attention Pooling obtained the highest ShanghaiTech AP (0.1060 ± 0.0073). Thus, the no pooling strategy dominated across domains and evaluation metrics. A controlled factorial comparison of Kernel MMD, CORAL, and DANN across all three pooling mechanisms showed that CORAL with Mean Pooling produced only a small numerical AUC increase to 0.5775 ± 0.0099, while Kernel MMD and several DANN configurations exhibited substantial seed variability. None of the selected pairwise comparisons was statistically significant after Holm correction. The findings should therefore be interpreted as setting‐specific tendencies: stronger source‐domain performance and the addition of global distribution‐alignment objectives did not consistently translate into better target‐domain anomaly localization in the examined UCF‐Crime‐to‐ShanghaiTech setting.
Zitieren
APA
Zawahra, M. T. A., Al-Batah, M. S., & Alzboon, M. S. (2026). An Empirical Analysis of MIL Pooling Under UCF‐Crime‐to‐ShanghaiTech Transfer in Weakly Supervised Video Anomaly Detection. Applied Computational Intelligence and Soft Computing, 2026(1). https://doi.org/10.1155/acis/1607118
Vancouver
Zawahra MTA, Al-Batah MS, Alzboon MS. An Empirical Analysis of MIL Pooling Under UCF‐Crime‐to‐ShanghaiTech Transfer in Weakly Supervised Video Anomaly Detection. Applied Computational Intelligence and Soft Computing. 2026;2026(1). doi:10.1155/acis/1607118
BibTeX
@article{Zawahra2026,
author = {Zawahra, Marwan Tayseer Al and Al-Batah, Mohammad Subhi and Alzboon, Mowafaq Salem},
title = {An Empirical Analysis of MIL Pooling Under UCF‐Crime‐to‐ShanghaiTech Transfer in Weakly Supervised Video Anomaly Detection},
journal = {Applied Computational Intelligence and Soft Computing},
year = {2026},
volume = {2026},
number = {1},
doi = {10.1155/acis/1607118}
}
EUVIMED