Estimating Depth from RGB and Sparse Sensing
This enables efficient transformation of sparse depth data from low-power sensors or SLAM systems into high-quality dense maps, with incremental improvements over existing sparse-to-dense methods.
The paper tackles the problem of generating dense depth maps from RGB images with very sparse depth measurements, achieving state-of-the-art results on NYUv2 and KITTI datasets with near real-time speeds and mean absolute error under 1% of actual depth for indoor scenes using only 1/256 of image pixels.
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.