Superpixel Cost Volume Excitation for Stereo Matching

TL;DR

The proposed approach capitalizes on the observation that neighboring pixels are predisposed to belong to the same object and exhibit closely similar intensities within the probability volume of superpixels, to encourage the network to generate consistent probability distributions of disparity within each superpixel.

Abstract

In this work, we concentrate on exciting the intrinsic local consistency of stereo matching through the incorporation of superpixel soft constraints, with the objective of mitigating inaccuracies at the boundaries of predicted disparity maps. Our approach capitalizes on the observation that neighboring pixels are predisposed to belong to the same object and exhibit closely similar intensities within the probability volume of superpixels. By incorporating this insight, our method encourages the network to generate consistent probability distributions of disparity within each superpixel, aiming to improve the overall accuracy and coherence of predicted disparity maps. Experimental evalua tions on widely-used datasets validate the efficacy of our proposed approach, demonstrating its ability to assist cost volume-based matching networks in restoring competitive performance.

Figures (7)

• Figure 1: Visualization of the real output distribution at boundaries on Scene Flow dataset. (a) is the input image, and its partial enlargement. (b) represents the disparity probability distribution of the superpixel belonging to the brown region. (c) and (d) show the output probability distributions of a given pixel from GwcNet and GwcNet$+$Ours. Our proposed methods rectify the incorrect distributions and avoid smoothness bias. Please zoom in to see the details.
• Figure 2: The proposed stereo matching framework consists of a stereo matching pipeline and a sub-network for superpixel segmentation. The superpixel branch (cyan) takes the left image as input and assists the stereo branch (black).
• Figure 3: Superpixel guided channel excitation module. The multi-scale short connections achieved through 2D convolutional kernels of varying sizes and strides combined with upsampling result in rich object context within the superpixel branch.
• Figure 4: The main components of the joint learning training head, which combines the output results from two branches, consist of a variance estimator for predicting matchbility and a superpixel pooling module, all driven by the cross-entropy loss function.
• Figure 5: Qualitative comparisons of ablation study on Scene Flow test set.