SETJiP: Spatial and Extra Temporal Jigsaw Puzzles for Video Anomaly Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video Anomaly Detection (VAD) is crucial for identifying unusual events in videos.
• Current VAD methods often use self-supervised learning, like Decoupled Spatial and Temporal Jigsaw Puzzles (DSTJiP).
• DSTJiP struggles with effectively supervising global motion examples, impacting VAD performance.

Purpose of the Study:

• To improve Video Anomaly Detection (VAD) performance by addressing limitations in self-supervised learning.
• To develop a method that provides more effective temporal supervision for global motion patterns.
• To enhance the robustness of VAD across diverse training data distributions.

Main Methods:

• Proposing Spatial and Extra Temporal Jigsaw Puzzles (SETJiP), an RGB-only self-supervised learning approach.
• Implementing two training schemes: one with additional temporal jigsaw puzzles and another with upweighted temporal jigsaw puzzles for global motion.
• Focusing on learning discriminative representations by predicting patch order.

Main Results:

• Both SETJiP training schemes demonstrated significant improvements over the original DSTJiP method.
• The proposed methods achieved competitive results compared to state-of-the-art VAD techniques.
• SETJiP effectively balances temporal supervision for global motion examples without performance degradation.

Conclusions:

• SETJiP offers a more effective approach to self-supervised video anomaly detection.
• The enhanced temporal supervision strategies are crucial for handling global motion effectively.
• SETJiP provides a robust and competitive solution for VAD benchmarks.