Abstract
Stereo disparity estimation is a critical task in computer vision, particularly for autonomous driving. While modern deep learning models utilize 4D cost volumes at reduced resolutions to manage computational costs, recovering high-resolution disparity maps remains a challenge. Traditional methods like linear interpolation often lead to blurry depth boundaries. This paper introduces Multistage Full Matching (MFM), a novel framework that decomposes the complex task of high-resolution similarity estimation into manageable stages. By decoupling similarity features and implementing a "Stages Mutual Aid" (SMA) strategy, MFM achieves state-of-the-art accuracy on major benchmarks, including Scene Flow and KITTI.
Introduction
Stereo matching involves finding the horizontal shift (disparity) of pixels between left and right image views to infer depth. To handle high-resolution imagery efficiently, most architectures construct a 4D cost volume at a lower resolution. However, the transition back to high-resolution disparity maps is often problematic. Linear interpolation assumes a smoothness in similarity scores that does not exist at object boundaries, while transposed convolutions suffer from "internal competition," where a single low-resolution feature must simultaneously predict multiple high-resolution scores.
Methodology
The proposed Multistage Full Matching (MFM) framework addresses these limitations through a "divide and conquer" strategy.
Task Decomposition and Feature Decoupling
MFM divides the matching range into $k$ cells and estimates similarity scores across $n$ distinct stages. This decoupling allows the network to specialize its weights for specific subsets of candidates, alleviating the bottleneck where one feature maps to many disparate points.
Stages Mutual Aid (SMA)
To prevent earlier stages from propagating errors, the SMA mechanism allows for collaborative learning. It merges outputs from previous stages to provide a reference distribution for the current stage. This "voting" mechanism acts as a regularizer, ensuring consistency across the pipeline.
Sub-pixel Refinement
The final high-resolution 3D cost volume undergoes sub-pixel refinement using a modified parabolic fitting method, ensuring high precision in the resulting disparity map.
Experiments and Results
The MFM model was evaluated on the Scene Flow, KITTI 2012, and KITTI 2015 benchmarks.
Quantitative Performance
On the Scene Flow dataset, MFM achieved an End-Point Error (EPE) of 0.66, significantly outperforming GwcNet (0.77). On KITTI 2015, it reduced the D1-all error to 1.87%, surpassing competitive models like AcfNet (1.89%). Ablation studies confirmed that the SMA module alone improved the EPE from 0.74 to 0.66.
Qualitative Analysis
Qualitative results demonstrate that MFM produces sharper depth boundaries and fewer artifacts compared to existing state-of-the-art methods.
Conclusion
The MFM framework demonstrates that high-resolution matching information can be effectively recovered from low-resolution features by treating the task as a learned, multistage decoupling process. By avoiding the pitfalls of linear interpolation and internal competition, MFM sets a new standard for accuracy in stereo disparity estimation. Future work may extend this multistage logic to other volumetric tasks such as optical flow or 3D object detection.
Summary updated