Benchmarking Compact VLMs for Clip-Level Surveillance Anomaly Detection Under Weak Supervision

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• CCTV safety monitoring requires anomaly detection with high accuracy and low latency.
• Weak supervision presents challenges for traditional anomaly detection methods.
• Compact vision-language models (VLMs) are explored for their potential in this domain.

Purpose of the Study:

• To investigate the effectiveness of parameter-efficiently adapted compact VLMs for CCTV anomaly detection.
• To establish a unified evaluation protocol for comparing different VLM approaches and baselines.
• To assess the trade-off between detection accuracy and per-clip latency.

Main Methods:

• A unified evaluation protocol was developed, standardizing preprocessing, prompting, dataset splits, metrics, and runtime settings.
• Compact VLMs were adapted using parameter-efficient fine-tuning.
• Performance was compared against training-free VLM pipelines and weakly supervised baselines.
• Metrics included accuracy, precision, recall, F1, ROC-AUC, and average per-clip latency.

Main Results:

• Parameter-efficiently adapted compact VLMs achieved performance comparable to or exceeding established methods.
• These models maintained competitive per-clip latency, crucial for real-time monitoring.
• Adaptation reduced prompt sensitivity, leading to more consistent behavior.
• A favorable accuracy-efficiency trade-off was demonstrated.

Conclusions:

• Parameter-efficient fine-tuning enables compact VLMs to function as reliable clip-level anomaly detectors.
• Compact VLMs offer a practical and efficient solution for CCTV safety monitoring under weak supervision.
• The proposed evaluation protocol ensures transparency and consistency in assessing anomaly detection methods.