Confidence Inference for Focused Learning in Stereo Matching

TL;DR

This paper introduces an unsupervised confidence inference method for stereo matching that uses a probabilistic interpretation of the loss function, enabling focused learning on high-confidence pixels and improving model convergence.

Contribution

It proposes a novel dense confidence map for stereo matching that relaxes the i.i.d. assumption, enhancing confidence estimation and training effectiveness.

Findings

Confidence maps correlate with pixel confidence levels.

Focused learning improves convergence and reduces overfitting.

Method outperforms traditional approaches in confidence estimation.

Abstract

In this paper, we present confidence inference approachin an unsupervised way in stereo matching. Deep Neu-ral Networks (DNNs) have recently been achieving state-of-the-art performance. However, it is often hard to tellwhether the trained model was making sensible predictionsor just guessing at random. To address this problem, westart from a probabilistic interpretation of theL1loss usedin stereo matching, which inherently assumes an indepen-dent and identical (aka i.i.d.) Laplacian distribution. Weshow that with the newly introduced dense confidence map,the identical assumption is relaxed. Intuitively, the vari-ance in the Laplacian distribution is large for low confidentpixels while small for high-confidence pixels. In practice,the network learns toattenuatelow-confidence pixels (e.g.,noisy input, occlusions, featureless regions) andfocusonhigh-confidence pixels. Moreover, it can be observed fromexperiments that the focused learning is very helpful in find-ing a better convergence state of the trained model, reduc-ing over-fitting on a given dataset.