EUVIMEDEUVIMED entdecken
Crossref

Improving Temporal Localization in Vision-Language Video Anomaly Detection

Siddharth Shah, Narendrasinh Chauhan

Journal of Innovative Image Processing · 05.06.2026 · 8 (3) · 765-785

DOI: 10.36548/jiip.2026.3.002 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Innovative Image Processing, online erschienen am 05.06.2026. Verfasst von Siddharth Shah und 1 weiteren Autor:innen. Die Arbeit nennt 21 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The recent developments in the vision-language-based model framework for weakly supervised video anomaly detection (WSVAD) have significantly enhanced anomaly detection performance. The dual-branch framework consisting of two branches for performing binary classification and aligning textual descriptors with visual snippets, has been found efficient concerning anomaly detection. Nonetheless, a significant problem of temporal localization still persists. The existing solutions use a fixed value of top-k snippets without consideration for either short or long anomalies. Moreover, prediction inconsistency in terms of temporal localization with the presence of spikes in normal periods and gaps in anomalies is another issue that arises. The current solution is based on the VadCLIP framework and only modifies some specific aspects of it. First, Confidence-Adaptive MIL (CA-MIL) computes a per-video threshold from the score distribution, selecting fewer snippets when confidence lowers and more when an anomalous event has a larger time frame. Second, a temporal smoothness term penalizes abrupt score transitions between adjacent snippets. Third, two parallel scoring heads, one point-wise MLP, and one local-context convolution are fused through learned gating that accounts for disagreement. Lastly, at test time, Score-level Temporal Context Aggregation (STCA) smooths the final predictions using local averaging and global statistics. Cross-modal attention provides a small additional boost to AUC. In UCF-Crime, the average mAP between the 0.1–0.5 IoU thresholds increases from 6.68 to 9.37 (+40.3%), with mAP@0.5. XD-Violence sees an average increase in mAP from 24.70 to 31.63 (+28.1%). Detection performance is preserved (UCF-Crime AUC decreases by 0.10% from 88.02 to 87.92; XD-Violence AP increases by 0.22% from 84.51 to 84.73).

Zitieren

APA

Shah, S., & Chauhan, N. (2026). Improving Temporal Localization in Vision-Language Video Anomaly Detection. Journal of Innovative Image Processing, 8(3), 765-785. https://doi.org/10.36548/jiip.2026.3.002

Vancouver

Shah S, Chauhan N. Improving Temporal Localization in Vision-Language Video Anomaly Detection. Journal of Innovative Image Processing. 2026;8(3):765-785. doi:10.36548/jiip.2026.3.002

BibTeX
@article{Shah2026,
  author = {Shah, Siddharth and Chauhan, Narendrasinh},
  title = {Improving Temporal Localization in Vision-Language Video Anomaly Detection},
  journal = {Journal of Innovative Image Processing},
  year = {2026},
  volume = {8},
  number = {3},
  pages = {765--785},
  doi = {10.36548/jiip.2026.3.002}
}
Kostenlos starten →