As an expression of three-dimensional information, depth map provides more possibilities for many computer vision applications, which puts forward higher quality requirements for depth maps. However, the spatial resolution of depth map is always low, which limits its potential. In this paper, we present a depth map super-resolution based on joint convolutional neural network (J-CNN). The network combines RGB image to guide the reconstruction of low-resolution depth map. Our model consists of three subnets. Subnet 1 uses simple convolutional layers to extract rough features of depth maps. Subnet 2 and 3 use a progressive up and progressive down sampling network structure. This structure can greatly expand the receptive fields and help to extract more fine features of the images at different scales. We use it to extract the complex features in the depth map and RGB image. Finally, convolutional layers connect three subnets to transfer useful information from RGB image to depth map. The final up-sampling reconstructed operation is realized by sub-pixel convolution, effectively avoiding the “checkerboard effect”. Our J-CNN is evaluated on Middlebury datasets which shows improved performance compared with six advanced methods. Our model also shows strong robustness under large sampling factor (16 times).
Paper
Full text
Color-guided depth map super resolution using joint convolutional neural network
Semantic Scholar · Computer Science · 2020
Abstract
As an expression of three-dimensional information, depth map provides more possibilities for many computer vision applications, which puts forward higher quality requirements for depth maps. However, the spatial resolution of depth map is always low, which limits its potential. In this paper, we present a depth map super-resolution based on joint convolutional neural network (J-CNN). The network combines RGB image to guide the reconstruction of low-resolution depth map. Our model consists of three subnets. Subnet 1 uses simple convolutional layers to extract rough features of depth maps. Subnet 2 and 3 use a progressive up and progressive down sampling network structure. This structure can greatly expand the receptive fields and help to extract more fine features of the images at different scales. We use it to extract the complex features in the depth map and RGB image. Finally, convolutional layers connect three subnets to transfer useful information from RGB image to depth map. The final up-sampling reconstructed operation is realized by sub-pixel convolution, effectively avoiding the “checkerboard effect”. Our J-CNN is evaluated on Middlebury datasets which shows improved performance compared with six advanced methods. Our model also shows strong robustness under large sampling factor (16 times).