One-Shot Learning for Surveillance Anomaly Recognition Using Siamese 3D CNN
Amin Ullah, Khan Muhammad, Killichbek Haydarov, Ijaz Ul Haq, Miyoung Lee, Sung Wook Baik*
Key Figure

  • Access the paper
    • Proceedings of the International Joint Conference on Neural Networks (IJCNN), 2020 published [📃 Full-Text]
  • Additional Link
    •   IEEE Xplore
    • Abstract

      • One-shot image recognition has been widely explored in computer vision; however, its application to video analytics remains limited. Surveillance anomaly recognition is a challenging problem, primarily due to the scarcity of accurately temporally annotated data. This paper addresses this limitation by introducing a one-shot learning framework for anomaly recognition. The proposed approach utilizes a Siamese 3D convolutional neural network to measure similarity between anomaly sequences. Additionally, the study evaluates existing 3D CNN architectures and proposes a lightweight 3D CNN model optimized for efficient one-shot anomaly recognition. Once trained, the model leverages discriminative 3D CNN features to detect anomalies in both new data and previously unseen classes. The framework is trained using the temporally annotated test set of the UCF Crime dataset and is further used to generate temporal labels for weakly annotated video data. Experimental results demonstrate the effectiveness of the proposed method in surveillance anomaly detection tasks.

    • Additional Comments