FBA-AMNET: Foreground-Background Aware Atrous Multiscale Networks for Stereo Disparity Estimation

In this paper, we propose novel networks for stereo disparity estimation. First, deep features are extracted using efficient depthwise-separable convolutions. Next, the stereo matching costs are calculated from the deep features with a novel extended cost volume. Then, rich multiscale contextual information is aggregated with the proposed atrous multiscale network (AMNet). The proposed foreground-background aware network (FBA-AMNET) is trained with an iterative multi-task learning strategy to discriminate between foreground and background objects at multiple scales. The proposed networks advance the state of the art on challenging disparity estimation benchmarks, such as the KITTI 2012, KITTI 2015, Sceneflow, and Middlebury stereo benchmarks.