This thesis investigates fine-tuning deep face features in a self-supervised manner for discriminative face representation learning, wherein we develop methods to automatically generate pseudo-labels for training a neural network. Most importantly solving this problem helps us to advance the state-of-the-art in representation learning and can be beneficial to a variety of practical downstream tasks. Fortunately, there is a vast amount of videos on the internet that can be used by machines to learn an effective representation. We present methods that can learn a strong face representation from large-scale data be the form of images or video.
However, while learning a good representation using a deep learning algorithm requires a large-scale dataset with manually curated labels, we propose self-supervised approaches to generate pseudo-labels utilizing the temporal structure of the video data and similarity constraints to get supervision from the data itself.
We aim to learn a representation that exhibits small distances between samples from the same person, and large inter-person distances in feature space. Using metric learning one could achieve that as it is comprised of a pull-term, pulling data points from the same class closer, and a push-term, pushing data points from a different class further away. ... mehrMetric learning for improving feature quality is useful but requires some form of external supervision to provide labels for the same or different pairs. In the case of face clustering in TV series, we may obtain this supervision from tracks and other cues. The tracking acts as a form of high precision clustering (grouping detections within a shot) and is used to automatically generate positive and negative pairs of face images. Inspired from that we propose two variants of discriminative approaches: Track-supervised Siamese network (TSiam) and Self-supervised Siamese network (SSiam). In TSiam, we utilize the tracking supervision to obtain the pair, additional we include negative training pairs for singleton tracks -- tracks that are not temporally co-occurring. As supervision from tracking may not always be available, to enable the use of metric learning without any supervision we propose an effective approach SSiam that can generate the required pairs automatically during training. In SSiam, we leverage dynamic generation of positive and negative pairs based on sorting distances (i.e. ranking) on a subset of frames and do not have to only rely on video/track based supervision.
Next, we present a method namely Clustering-based Contrastive Learning (CCL), a new clustering-based representation learning approach that utilizes automatically discovered partitions obtained from a clustering algorithm (FINCH) as weak supervision along with inherent video constraints to learn discriminative face features. As annotating datasets is costly and difficult, using label-free and weak supervision obtained from a clustering algorithm as a proxy learning task is promising. Through our analysis, we show that creating positive and negative training pairs using clustering predictions help to improve the performance for video face clustering.
We then propose a method face grouping on graphs (FGG), a method for unsupervised fine-tuning of deep face feature representations. We utilize a graph structure with positive and negative edges over a set of face-tracks based on their temporal structure of the video data and similarity-based constraints. Using graph neural networks, the features communicate over the edges allowing each track's feature to exchange information with its neighbors, and thus push each representation in a direction in feature space that groups all representations of the same person together and separates representations of a different person.
Having developed these methods to generate weak-labels for face representation learning, next we propose to learn compact yet effective representation for describing face tracks in videos into compact descriptors, that can complement previous methods towards learning a more powerful face representation. Specifically, we propose Temporal Compact Bilinear Pooling (TCBP) to encode the temporal segments in videos into a compact descriptor. TCBP possesses the ability to capture interactions between each element of the feature representation with one-another over a long-range temporal context. We integrated our previous methods TSiam, SSiam and CCL with TCBP and demonstrated that TCBP has excellent capabilities in learning a strong face representation. We further show TCBP has exceptional transfer abilities to applications such as multimodal video clip representation that jointly encodes images, audio, video and text, and video classification.
All of these contributions are demonstrated on benchmark video clustering datasets: The Big Bang Theory, Buffy the Vampire Slayer and Harry Potter 1. We provide extensive evaluations on these datasets achieving a significant boost in performance over the base features, and in comparison to the state-of-the-art results.