Relative Depth Estimation from Hyperspectral Data

Abstract

Listen

Translate article

This paper addresses the problem of relative depth estimation using spatial defocus and spectral chromatic aberration presented in hyperspectral data. Our approach produces merged relative sparse depth map using two different methods. The first method constructs a histogram descriptor for edge pixels in each spectral band image. Due to the spectral chromatic aberration, different edge statistical information can be extracted from each band even at the same location. Variance among histogram bins provides input data for band-wise spatial defocus calculation. These band-wise statistical data are later combined to give the first sparse depth map. The second approach uses difference of neighboring spectral vectors to estimate relative depth. The two sparse maps with distinguishing features are finally combined and optimized to generate final sparse depth map. During the last step, normalization and smoothing are used to guarantee better consistency among edge pixels. Experimental results show that our method can generate better sparse depth map than alternative methods which operate on RGB images.