An $\ell_{1/2}$ -Norm Regularizer-Based Sparse Coding Framework for Gaze Prediction in First-Person Videos

Abstract

Predicting human gaze is important for efficiently processing and understanding numerous incoming visual information from first-person videos (FPVs). Even though people continuously gaze in noisy environments, most existing gaze prediction algorithms are based on saliency mapping, which is sensitive to noisy surroundings in the real world. Sparsity-based saliency detection algorithms perform favorably against state-of-the-art methods. In this paper, we apply a novel saliency detection method based on sparse coding with the l 1/2 -norm for predicting human gaze in FPVs. Image boundaries are first extracted via superpixels as bases for a dictionary, from which a sparse representation model is constructed. For each superpixel, we first compute sparse reconstruction errors. Then, a saliency map is updated based on the reconstruction errors. To receive the sparse reconstruction errors, the most widely utilized sparse constraint is the l 1 -norm. However, the l 1 -norm leads to over-penalization of large components in a sparse vector. We employ the l 1/2 -norm for sparse coding, which can lead to a sparser solution for a more accurate gaze prediction than the l 1 -norm. We transform the complex nonconvex optimization of sparse coding with the l 1/2 -norm to a number of one-dimensional minimization problems. In this way, we obtain the closed-form solutions efficiently. The experimental results using a real-world gaze dataset demonstrate that the proposed algorithm performs better than the state-of-the-art methods of gaze prediction for FPVs.