Lighter and Faster Two-Pathway CMRNet for Video Saliency Prediction

Abstract

Existing dynamic saliency prediction models face challenges like inefficient spatio-temporal feature integration, ineffective multi-scale feature extraction, and lacking domain adaptation because of huge pre-trained backbone networks. In this paper, we propose a two pathway architecture with effective feature integration of spatial and temporal domains at multiple scales for video saliency prediction. Frame and optical flow pathways extract features from video frame and optical flow maps, respectively using a series of cross-concatenated multi-scale residual (CMR) blocks. We name this network as two-pathway CMRNet (TP-CMRNet). Every CMR block follows a feature fusion and attention module for merging features from two pathways and guiding the network to weigh salient regions, respectively. A bi-directional LSTM module is used for learning the task by looking at previous and next video frames. We build a simple decoder for feature reconstruction into the final attention map. TP-CMRNet is comprehensively evaluated using three benchmark datasets: DHF1K, Hollywood-2, and UCF sports. We observe that our model performs at par with other deep dynamic models. In particular, we outperform all the other models with a lesser number of model parameters and lower inference time.