Towards a sparse low-rank regression model for memorability prediction of images.

Nowadays, it is inevitable to experience plenty of images in everyday life. Some of them are remembered for a long time while others are forgotten after only a glance. It has been proved that memorability is an intrinsically stable property of images which measures the degree to which images are remembered. Although some work have been conducted to investigate the factors that make an image memorable, yet studies on designing robust models to predict image memorability have rarely been reported. Inspired by the good property of Low-Rank Representation (LRR) in dealing with noisy data, in this paper we propose a sparse low-rank regression framework for image memorability prediction, in which a projection matrix, applied to capture the global low-rank structure embedded in original feature space, and a sparse coefficient vector, applied to build connections between images and their memorability scores, are jointly learnt to guarantee the superior performance. In particular, to enable our proposed approach to discover discriminant attribute features automatically, we impose a structured sparsity constraint on the reconstruction error matrix against the existence of noisy attributes. We develop an alternating direction algorithm by applying augmented Lagrangian multipliers method to solve the objective function of our model. Experiments conducted on two publicly available memorability datasets demonstrates the effectiveness of the proposed method. Source code is freely available: https://www.github.com/HodorHoldthedoor/image-memorability.