Subject

Video anomaly detection refers to the task of detecting unexpected events or behaviours in a video sequence. In recent years, many deep learning models were proposed to solve this task, each with a certain degree of success. These include convolutional neural networks (CNNs), which learn a set of discriminative features to classify video frames as anomalous or not. However, the anomaly detection task comes with a series of challenges: (1) anomalies are unpredictable in essence and do not occur frequently thus, deep models may generalize poorly to new abnormalities due to the lack of training examples. (2) How deep models come to a decision may be opaque, thus visual explanations must be produced to understand the decision making of these models. In practice, such explanations can consist of “saliency maps” which highlight the specific objects in the scene which trigger the positive response of the classification model. Therefore, in this thesis, we propose to investigate a variety of model explanation techniques, applied to recent anomaly detection models.

Kind of work

The student(s) will be responsible for implementing and evaluating deep learning models on video datasets and implement model explanation techniques for these models. A comparative study is expected to be conducted.

Framework of the Thesis

Sultani, W., Chen, C., & Shah, M. (2018). Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 6479-6488).
Montavon, Grégoire, et al. "Explaining nonlinear classification decisions with deep taylor decomposition." Pattern Recognition 65 (2017): 211-222.

Number of Students

1-2

Expected Student Profile

The student should have a good background in linear algebra, machine learning and image processing.
The student must be proficient in Python and should be able to learn deep learning techniques from existing codes and documentation. Knowledge of deep learning and libraries like PyTorch, Keras or TensorFlow is a plus.