The Sampling-Gaussian for stereo matching

TL;DR

A novel method called Sampling-Gaussian is proposed as a substitute for soft-argmax that improves accuracy without increasing inference time and achieves significant improvements on small datasets and models with weaker generalization capabilities.

Abstract

The soft-argmax operation is widely adopted in neural network-based stereo matching methods to enable differentiable regression of disparity. However, network trained with soft-argmax is prone to being multimodal due to absence of explicit constraint to the shape of the probability distribution. Previous methods leverages Laplacian distribution and cross-entropy for training but failed to effectively improve the accuracy and even compromises the efficiency of the network. In this paper, we conduct a detailed analysis of the previous distribution-based methods and propose a novel supervision method for stereo matching, Sampling-Gaussian. We sample from the Gaussian distribution for supervision. Moreover, we interpret the training as minimizing the distance in vector space and propose a combined loss of L1 loss and cosine similarity loss. Additionally, we leveraged bilinear interpolation to upsample the cost volume. Our method can be directly applied to any soft-argmax-based stereo matching method without a reduction in efficiency. We have conducted comprehensive experiments to demonstrate the superior performance of our Sampling-Gaussian. The experimental results prove that we have achieved better accuracy on five baseline methods and two datasets. Our method is easy to implement, and the code is available online.

Figures (12)

• Figure 1: Quantitative comparisons on Sceneflow and Kitti. We implement our Sampling-Gaussian (SG) with five baseline methods for comparison. They are MSN2D and MSN3D shamsafar_mobilestereonet_2021, PSMnetchang_pyramid_2018, GwcNet-gguo_group-wise_2019, IGEV-Stereoxu2023iterative
• Figure 2: The left plot shows the impact of the disparity range. The predicted disparity near the endpoints is deviated. The right plot shows that after trilinear interpolation, the probabilities are linearly distributed and are unable to fit the Gaussian distribution.
• Figure 3: The workflow of our proposed Sampling-Gaussian
• Figure 4: The left plot: Since the probabilities are not linearly related, therefore it can fit into the Gaussian distribution. The middle plot: The loss landscape of L1 loss, dashed lines are contour line. Two vectors with the same L1 loss, could result in significant difference in EPE. The right plot: The loss landscape of the combined loss and the EPE of two vectors with the same loss are similar.
• Figure 5: Qualitative comparisons on Sceneflow