Learning-Deep-Learning

Deep Optics for Monocular Depth Estimation and 3D Object Detection

October 2019

tl;dr: End-to-end design of optics and imaging process using coded defocus as additional depth cue.

Overall impression

This paper introduces the idea of adding a defocus blur and aberration without hurting the 2D performance. The idea of co-designing optics and image processing is core to computation photography.

That means using a special lens we can capture a carefully “blurred” image from which the depth information can be recovered easily for depth-dependent task such as monocular 3D object detection.

This paper proves the feasibility of an interesting idea, but it has a long way to go before industrial application. In addition, how this would impact the detection of small object is yet to be proved.

Key ideas

Technical details

We do not claim to conclusively surpass existing methods, as we use the ground truth or pseudo-truth depth map in simulating our sensor images, and we are limited to an approximate, discretized, layer-based image formation model.

Lens optimized for depth estimation maintains 2D object detection performance while further improving 3D object detection from a single image

Notes