Deep Feature Filtering and Contextual Information Gathering Network for RGB-D Salient Object Detection

Along with the update of depth sensors, salient object detection (SOD) models are evolving in dividend and are being increasingly investigated. Due to the essential difference between depth and RGB images, existing SOD models mostly focus on finding a reasonable way to facilitate multimodal (i.e., RGB and depth) features fusion while ignoring the interference of low-resolution depth maps on the inference stage. To this end, we first propose a Deep Feature Filtering and Contextual Information Gathering Network for RGB-D Salient Object Detection (DCNet). First, a filter based on an attention mechanism is used to select important information in the depth features, which can adequately filter the low-quality depth maps. Moreover, before integrating multi-level feature complementarity, we incorporate a pre-processing step, the contextual information gathering module, which consists of separable (Sep) convolution, dilated (Dil) convolution, adaptive averaging pooling (AVP), and residual networks, to capture information diversity from a global perspective. We conclude by proposing a new mixed loss function based on binary cross-entropy loss (BCE) and intersection over union loss (IOU) for supervised network optimization. With four evaluation metrics on five public datasets, DCNet outperformed ten start-of-the-art methods in terms of simplicity and efficiency.