Pseudo-Label Generation and Refinement for Minimally Supervised Video Anomaly Detection

Document Type

Dissertation

Abstract

Anomaly detection is a critical research area with significant real-world applications. However, given the rare occurrences of anomalies, learning to localize anomalies in long videos can be incredibly challenging. Additionally, annotating long videos is time-consuming and laborious. Therefore, weakly supervised methods have gained popularity in recent years. However, the widely used Multiple Instance Learning (MIL) approaches in weakly supervised anomaly detection have several limitations, including being impacted by label noise, requiring videos to have an identical number of segments, and having a complete video input at each training iteration. To mitigate these issues, we propose a novel training framework in which a training batch is formed using random features extracted from different videos in a dataset, minimizing the correlation between feature vectors in a given training batch. We also introduce pseudo-label generation and refinement to create feature-vector-level pseudo-labels iteratively. We take the pseudo-labeling up a notch by introducing an unsupervised clustering algorithm to generate meta-labels without utilizing any labels, thus facilitating fully-unsupervised training. Our weakly labeled and unsupervised approach variants outperform several existing state-of-the-art methods while reducing the training complexity.

Publication Date

6-2023

Comments

Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfillment of the requirements for the M.Sc degree in Computer Vision

Advisors: Dr. Karthik Nandakumar, Dr. Fahad Khan

with 1 year embargo period

This document is currently not available here.

Share

COinS