ABSTRACT State‐of‐the‐art stereo matching models perform well on specific datasets, but data uncertainty limits their stability in unknown scenarios. In this paper, we propose VMI‐Stereo, a variational Bayesian inference based multivariate information fusion network, to enhance robustness and generalization. It designs a variational Bayesian approach for feature extraction and cost aggregation to data uncertainty and improve generalization to distribution shifts. Then, we introduce Fourier gradient perturbations to capture edge and texture information, strengthening uncertainty characterization. These advantages are integrated via the Fourier gradient perturbation volume, which fuses Fourier gradient and variational Bayesian features to enhance generalization to data uncertainty. Experiments show that trained on Scene Flow, our model reduces EPE by 8.38% over the baseline and achieves superior generalization on KITTI and ETH3D. These results confirm that VMI‐Stereo has significant advantages in handling stereo matching tasks in complex, uncertain environments.
Zhao et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: