Multilingual-Prompt-Guided Directional Feature Learning for Weakly Supervised Video Anomaly Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised video anomaly detection uses video-level labels for cost-efficient annotation.
• Challenges include diverse/incomplete anomalies and difficulties in prompt design for vision-language models.
• Existing methods struggle with real-world scenario diversity and annotation workload.

Purpose of the Study:

• To improve feature learning in weakly supervised video anomaly detection.
• To address challenges in prompt design for diverse real-world scenarios.
• To enhance the accuracy and efficiency of anomaly detection systems.

Main Methods:

• Integration of multilingualism and multiple prompts for enhanced feature learning.
• Adaptive top-K prompt selection across different linguistic domains.
• A multi-granularity attention module combining Transformer and Mamba for visual feature enhancement.
• Incorporation of multilingual prompt guidance loss and gradual directional loss.

Main Results:

• Demonstrated effectiveness on four diverse video anomaly detection datasets.
• Showcased generalizability on two medical datasets (EMG and ECG temporal data).
• Improved feature learning and anomaly detection performance through proposed methods.

Conclusions:

• The proposed method effectively enhances weakly supervised video anomaly detection.
• Multilingual prompts and advanced attention mechanisms offer a robust solution.
• The approach shows promise for both general and specialized (medical) anomaly detection tasks.