Analysis of KITTI Data for Stereo Analysis with Stereo Confidence Measures

Abstract

Abstract. The recently published KITTI stereo dataset provides a new quality of stereo imagery with partial ground truth for benchmarking stereo matchers. Our aim is to test the value of stereo confidence measures (e.g. a left-right consistency check of disparity maps, or an analysis of the slope of a local interpolation of the cost function at the taken minimum) when applied to recorded datasets, such as published with KITTI. We choose popular measures as available in the stereo-analysis literature, and discuss a naive combination of these. Evaluations are carried out using a sparsification strategy. While the best single confidence measure proved to be the right-left consistency check for high disparity map densities, the best overall performance is achieved with the proposed naive measure combination. We argue that there is still demand for more challenging datasets and more comprehensive ground truth.

Ensemble Learning for Confidence Measures in Stereo Vision

“…Apart from bias that may have been introduced due to flaws in the ground truth data [10] used here, advantages of the proposed method are larger where stereo is more challenging and hence produces more error prone results. Yet, to shed light on this, new challenges for stereo need to be defined (and come with ground truth), beyond what is present in KITTI data.…”

Learning to Detect Ground Control Points for Improving the Accuracy of Stereo Matching

“…Other recent research on confidence estimation, from which we draw inspiration and borrow features, includes the work of Reynolds et al [20] on time-offlight data and of Hu and Mordohai [9] on stereo. Haeusler and Klette [6] also considered several confidence measures, as well as the product of all measures, demonstrating good performance in sparsification. Pfeiffer et al [19] integrated three confidence measures into a mid-level representation for 3D reconstruction and showed that Bayesian reasoning outperforms sparsification by thresholding.…”

“…What separates our approach from recent literature on confidence estimation [20,6,9,21,7], regardless of the use of learning, is that the main objective of these methods is sparsification. They can indeed generate disparity maps with progressively fewer errors by removing matches starting from the least reliable ones.…”

Ensemble Classifier for Combining Stereo Matching Algorithms

“…Relevant to our research is the literature on confidence estimation [18,13,28] and on learning optimization or regularization parameters [41,37] for stereo. The latter methods aim to learn a small number of global parameters, such as the weights of the data and smoothness terms of an MRF, while our work aims to train classifiers that make a decision per pixel based on local features and context.…”