Bottom-up Fixation Prediction Using Unsupervised Hierarchical Models

Fixation prediction, also known as saliency modelling, has been a subject undergoing intense study in various contexts. In the context of assistive vision technologies, saliency modelling can be used for development of simulated prosthetic vision as part of the saliency-based cueing algorithms. In this paper, we present an unsupervised multi-scale hierarchical saliency model, which utilizes both local and global saliency pipelines. Motivated by bio-inspired vision findings, we employ features from image statistics. Contrary to previous research, which utilizes one-layer equivalent networks such as independent component analysis (ICA) or principle component analysis (PCA), we adopt independent subspace analysis (ISA), which is equivalent to a two-layer neural architecture. The advantage of ISA over ICA and PCA is robustness towards translation meanwhile being selective to frequency and rotation. We extended the ISA networks by stacking them together, as done in deep models, in order to obtain a hierarchical representation. Making a long story short, (1) we define a framework for unsupervised fixation prediction, exploiting local and global saliency concept which easily generalizes to a hierarchy of any depth. (2) we assess the usefulness of the hierarchical unsupervised features, (3) we adapt the framework for exploiting the features provided by pre-trained deep neural networks, (4) we compare the performance of different features and existing fixation prediction models on MIT1003, (5) we provide the benchmark results of our model on MIT300.