BROWSE
Detail
Conference Paper
Fighting Against Data Deficiency Problems for Surveillance
Cited - time in 
Share 
Share
- Authors
- Seung-Ik Lee, Jinha-Lee, Marcella Astrid, Muhammad Zaigham Zaheer
- Issue Date
- 2021-10
- Citation
- International Conference on Control, Automation and Systems (ICCAS) 2021, pp.1-3
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Anomaly detection has recently become a popular domain the field of computer vision. Particularly, an upward trend in analysing pedestrian surveillance has been observed due to its popular demand and real-world applications. Typically, autoencoders are used to train on one-class (normal data) and at test time, the models are expected to produce high reconstruction errors for anomalous inputs, that correspond to high anomaly score. However, due to the unavailability of anomaly examples, the network often becomes too generalize and does not perform well. In this paper, we explore two different approaches, negative learning and pseudo-anomaly generation, to improve the anomaly classification capability of a model. Experiments are conducted using a real-world surveillance dataset recorded using the real CCTVs installed in a street. Experiments suggests that while negative learning is better due to the usage of real-anomaly examples, pseudoanomaly based method also provides comparable performance with an additional benefit of not utilizing real-anomaly examples.
- KSP Keywords
- Anomaly classification, Computer Vision(CV), Real-world applications, Test Time, anomaly detection, anomaly score