We present a deep model that can accurately produce dense depth maps given an RGB image with known depth at a very sparse set of pixels. The model works simultaneously for both indoor/outdoor scenes and produces state-of-the-art dense depth maps at nearly real-time speeds on both the NYUv2 and KITTI datasets. We surpass the state-of-the-art for monocular depth estimation even with depth values for only 1 out of every ~10000 image pixels, and we outperform other sparse-to-dense depth methods at all sparsity levels. With depth values for 1/256 of the image pixels, we achieve a mean absolute error of less than 1% of actual depth on indoor scenes, comparable to the performance of consumer-grade depth sensor hardware. Our experiments demonstrate that it would indeed be possible to efficiently transform sparse depth measurements obtained using e.g. lower-power depth sensors or SLAM systems into high-quality dense depth maps.
@article{arxiv.1804.02771,
title = {Estimating Depth from RGB and Sparse Sensing},
author = {Zhao Chen and Vijay Badrinarayanan and Gilad Drozdov and Andrew Rabinovich},
journal= {arXiv preprint arXiv:1804.02771},
year = {2018}
}
Comments
European Conference on Computer Vision (ECCV) 2018. Updated to camera-ready version with additional experiments