DenseLiDAR: A Real-Time Pseudo Dense Depth Guided Depth Completion Network

Depth Completion can produce a dense depth map from sparse input and provide more complete 3D description of the environment. Despite great progress made in completion, sparsity low density ground truth still make this problem challenging. In work, we propose DenseLiDAR, novel real-time pseudo-depth guided completion neural network. We exploit obtained simple morphological operations to guide network three aspects: (1) Constructing residual structure for output; (2) Rectifying data; (3) Providing structural loss training Thanks these designs, higher performance output could be achieved. addition, two new metrics better evaluating quality predicted are also presented. Extensive experiments on KITTI benchmark suggest that our model is able achieve state-of-the-art at highest frame rate 50 Hz. The further evaluated by several downstream robotic perception or positioning tasks. For task object detection, 3~5 percent gains small objects categories achieved detection dataset. RGB-D SLAM, accuracy vehicle's trajectory Odometry These promising results not only verify high prediction, but demonstrate potential improving related tasks using results.