Abstract:
Depth completion aims to recover a dense depth map given sparse depth samples and optional additional data as input. While some methods take only sparse data as input,
others consider the corresponding RGB image as guidance to get a better dense
depth representation. With the rise of data driven neural networks, most computer
vision researchers moved away from classical methods and exploited the power of
Convolutional neural Network (CNN) for recovering accurate and dense depths.
Some classical handcrafted methods also provide a commensurate result as that of
modern deep neural network methods. In this paper we first evaluate and compare different modern learning-based depth completion approaches and following that we devise an algorithm that Mix classical and learning-based methods. The proposed method combines the two approaches to take advantage of the two methods and can be trained from end-to-end. Then we evaluate our algorithm on the challenging KITTI depth completion benchmark and NYU-depth-v2 dataset.
