Search

Eye movements in the case of freely viewing natural scenes are believed to be guided by local contrast, global contrast, and top-down visual factors. Although a lot of previous works have explored these three saliency cues for several years, there still exists much room for improvement on how to model them and integrate them effectively. This paper proposes a novel computation model to predict eye fixations, which adopts a multiresolution convolutional neural network (Mr-CNN) to infer these three types of saliency cues from raw image data simultaneously. The proposed Mr-CNN is trained directly from fixation and nonfixation pixels with multiresolution input image regions with different contexts. It utilizes image pixels as inputs and eye fixation points as labels. Then, both the local and global contrasts are learned by fusing information in multiple contexts. Meanwhile, various top-down factors are learned in higher layers. Finally, optimal combination of top-down factors and bottom-up contrasts can be learned to predict eye fixations. The proposed approach significantly outperforms the state-of-the-art methods on several publically available benchmark databases, demonstrating the superiority of Mr-CNN. We also apply our method to the RGB-D image saliency detection problem. Through learning saliency cues induced by depth and RGB information on pixel level jointly and their interactions, our model achieves better performance on predicting eye fixations in RGB-D images. [ABSTRACT FROM AUTHOR]

In this paper, we present a unified representation based on the spatio-temporal steerable pyramid (STSP) for the holistic representation of human actions. A video sequence is viewed as a spatio-temporal volume preserving all the appearance and motion information of an action in it. By decomposing the spatio-temporal volumes into band-passed sub-volumes, the spatio-temporal Laplacian pyramid provides an effective technique for multi-scale analysis of video sequences, and spatio-temporal patterns with different scales could be well localized and captured. To efficiently explore the underlying local spatio-temporal orientation structures at multiple scales, a bank of three-dimensional separable steerable filters are conducted on each of the sub-volume from the Laplacian pyramid. The outputs of the quadrature pair of steerable filters are squared and summed to yield a more robust oriented energy representation. To be further invariant and compact, a spatio-temporal max pooling operation is performed between responses of the filtering at adjacent scales and over spatio-temporal neighbourhoods. In order to capture the appearance, local geometric structure and motion of an action, we apply the STSP on the intensity, 3D gradients and optical flow of video sequences, yielding a unified holistic representation of human actions. Taking advantage of multi-scale, multi-orientation analysis and feature pooling, STSP produces a compact but informative and invariant representation of human actions. We conduct extensive experiments on the KTH, UCF Sports and HMDB51 datasets, which shows the unified STSP achieves comparable results with the state-of-the-art methods. [ABSTRACT FROM AUTHOR]