We solve the problem of video object segmentation by investigating how to expand the role of convolution in convolutional neural networks. Based on the One-Shot Video Object Segmentation (OSVOS) which can successfully tackle the task of semi-supervised video object segmentation, we introduce U-shape architecture. We first build a Global Guidance Module (GGM) on the bottom-up path to provide location information of potentially significant objects for layers of different feature levels. Then we design a Multi-scale Convolution Module (MCM) to fully get feature information and a Feature Fusion Module (FFM) to make the coarse-level semantic information well fused with the finelevel features from the top-down pathway. GGM and FFM allow the high-level semantic features to be progressively refined, yielding detail enriched segmentation maps. The experimental results on DAVIS 2016 data set shows that our proposed approach can more accurately locate the segmentation objects with sharpened details and our model has improved on all indicators than OSVOS.
Saliency detection is a fundamental problem in computer vision and image processing. Deep learning models perform better than traditional approaches but suffer from their huge parameters and slow speeds. To handle these problems, in this paper we propose the Dense Dilation Network(DDN) -- a novel saliency model based on Dense Convolutional Network(DenseNet), all layers' feature maps are extracted in the same resolution through a step-wise upsampling procedure which contains dilated convolution filters and deconvolution filters. To derive saliency maps, we propose a fusion dense block which contains dilated convolution filters and 1×1 Conv layers to fuse all low-level and high-level feature maps. DDN is an end-to-end architecture which does not need any post-processing procedures. Experiments on the benchmark datasets demonstrate that DDN outperforms the state-of-the-art methods on saliency detection and it requires less parameters and computation time.