Attention Unet++ for lightweight depth estimation from sparse depth samples and a single RGB image

Depth estimation from a single RGB image with sparse depth measurements has already been proved to be an effective way of predicting dense and high-precision depth maps. However, most of its networks are based on comparatively complex and fixed architectures that are too slow and inflexible to pursue the maximum task performance for various conditions. Addressing this problem, we proposed a flexible and lightweight network architecture that can be split into a series of sub-networks with different accuracy and parameter size in use. We verified our proposed method’s effectiveness on NYUv2 and KITTI Odometry datasets. The results show that it is possible to achieve approximate accuracy as prior works but at the number of parameters that are an order of magnitude smaller. Our methodology’s most efficient sub-network performs the best for balancing the computation and accuracy with only no more than 1 M parameters.