PROJECT TITLE :

Hierarchical Features Driven Residual Learning for Depth Map Super-Resolution

ABSTRACT:

Computer vision activities such as intelligent cars and 3D reconstruction can be made easier by the rapid development of inexpensive and portable depth cameras for consumers. Because of this, the limited spatial resolution of low-cost depth sensors (e.g., the Kinect) restricts their usefulness. A new deep network for depth map super-resolution (SR), named DepthSR-Net, is proposed in this research. Hierarchical features driven residual learning is used to automatically infer a high-resolution (HR) depth map from its low-resolution (LR) version. U-Net deep network architecture is used in DepthSR-design. Net's Bicubic interpolation upsampling and input pyramid construction yield multiple level receptive fields from a depth map. Next, we extract hierarchical features from the input pyramid, intensity image, and encoder-decoder structure of U-Net, as shown in Figure 1. Finally, we use the hierarchical features to calculate the difference between the interpolated depth map and the HR one. The interpolated depth map is used to create the final HR depth map, which is then calculated using the learnt residual. Each component of the proposed network is tested using an ablation research. Research shows that the proposed method is superior to current methods. Other low-level visual problems could also benefit from the suggested network.