Open-set object recognition

Abstract

Despite significant advances in object recognition and classification over the past couple of decades, there are various situations where collecting representative training samples from all classes in real-world scenarios is quite expensive, or the system may be exposed to unpredictable novel samples at the test time. The pattern classification problem is commonly referred to as an open-set recognition task in such cases where limited and incomplete knowledge of the entire data distribution is provided to the model during the training time. During test phase, unknown classes can be faced which requires the classifier to accurately classify the previously seen classes while effectively rejecting unseen ones. Among others, one-class classification serves as a plausible solution to the open-set recognition problem. Nevertheless, current one-class classifiers have their limitations. Classical kernel-based approaches require carefully designed features to obtain reasonable performance but rest on a solid basis in statistical learning theory, providing good robustness against training set impurities. More recent deep learning-based methods, on the other hand, focus on learning relevant features directly from the data but typically rely on ad hoc one-class loss functions, which very often do not generalize well and are not robust against the omnipresent noise and contamination in the training set. In this thesis, we introduce a novel approach which leverages the advantages of both kernel-based and deep-learning approaches by bringing the two learning formalisms under a common umbrella. In particular, the proposed method learns deep convolutional features to optimize a kernel Fisher null-space loss subject to a Tikhonov regularisation on the discriminant in the Hilbert space. As such, it can be trained in a deep end-to-end fashion while being robust against training set contamination. Through extensive experiments conducted on different image datasets in various evaluation settings, the proposed approach is shown to be quite robust and more effective than the current state-of-the-art methods for anomaly detection in the scenario where the training set is corrupted and contains noisy samples. At the same time, the proposed approaches can be effectively utilized in an unsupervised scenario to rank the data points based on their conformity with the majority of samples.