Sparse Stereo Disparity Map Densification Using Hierarchical Image Segmentation

Abstract

We describe a novel method for propagating disparity values using hierarchical segmentation by waterfall and robust regression models. High confidence disparity values obtained by state of the art stereo matching algorithms are interpolated using a coarse to fine approach. We start from a coarse segmentation of the image and try to fit each region's disparities using robust regression models. If the fit is not satisfying, the process is repeated on a finer region's segmentation. Erroneous values in the initial sparse disparity maps are generally excluded, as we use robust regressions algorithms and left-right consistency checks. Final disparity maps are therefore not only denser but can also be more accurate. The proposed method is general and independent from the sparse disparity map generation: it can therefore be used as a post-processing step for any stereo-matching algorithm.

Continuous 3D Label Stereo Matching Using Local Expansion Moves

“…Since our model, especially our data term borrowed from [5], does not consider such new difficulties, we here slightly modify our model for adapting it to the latest benchmark. To this end, we incorporate the state-of-the-art CNN-based matching cost function by Zbontar and LeCun [49], by following manners of current top methods [10], [20], [28], [29] on the latest benchmark. Here, we only replace our pixelwise raw matching cost function ρ(·) of Eq.…”

“…Our method achieves the current best average error rate among 64 existing algorithms. We here list current top methods [10], [20], [28], [29], [49] that use the matching costs of MC-CNN-acrt [49] like ours. (Snapshot on July 4, 2017) except for bad 4.0 -all.…”

CBMV: A Coalesced Bidirectional Matching Volume for Disparity Estimation

“…CBMV disparity map Figure 1. The left view of Djembe stereo dataset [34] along with the disparity map computed by CBMV [2,6,8,15,18,37,38,43]. Our goal is similar to MC-CNN, since we also aim to estimate a matching volume that can be used as input to various optimization algorithms enabling them to produce highly accurate disparity maps.…”

“…Similar Siamese networks followed by the fast or accurate similarity estimation subnetworks have also been proposed by [4,11,20,49], while more recently, other authors have increased the effective receptive field of the networks without loss of resolution [26,39,47]. Many of the other top ranked methods have either been inspired by MC-CNN or directly use it to compute the matching cost [2,6,8,15,18,37,38,43].…”

Asymmetric Learning for Stereo Matching Cost Computation

“…siamese network with several traditional post-processing steps and obtained impressive results. Taking advantage of the CNN similarity computation [9], many advanced stereo methods have been proposed [10]- [12]. In contrast to the patch-based methods, end-to-end networks [13], [14] are proposed to estimate per-pixel disparity straightforwardly, without any post-processing or regularization.…”