Depth Completion Network Guided by RGB Attention

The goal of the depth completion task is to generate a dense depth map from a sparse depth map. The RGB image is usually introduced as a guide to help the network obtain the missing depth information. Due to the rich semantic features of RGB images, the method based on RGB guidance could obtain better performance. However, RGB guidance will bring wrong information in feature fusion, because the segmentation boundary of RGB features does not match the real depth distribution. A depth completion network based on the attention mechanism is proposed to correct the wrong fusion information. Bidirectional attention is introduced to correct wrong segmentation boundaries in RGB images. Compared with the SOTA methods, the proposed method still maintains excellent performance under different sparse input conditions. And the proposed method has shorter inference time and smaller model size, while obtaining competitive results.