Efficient Multi-Scale Stereo-Matching Network Using Adaptive Cost Volume Filtering.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• State-of-the-art stereo matching networks often use computationally intensive 3D convolutions for cost volume aggregation, limiting mobile deployment.
• Existing methods struggle with ambiguous regions due to indirect cost volume supervision, leading to overfitting.
• Previous attempts to regularize cost volumes with ground truth disparity are insufficient.

Purpose of the Study:

• To develop an efficient deep learning network for stereo matching suitable for resource-limited environments.
• To address limitations in cost volume aggregation and supervision for improved disparity estimation.
• To enhance real-time performance and accuracy in stereo matching tasks.

Main Methods:

• Proposed an efficient multi-scale sequential feature fusion network (MSFFNet) using parallel multi-scale SFF modules and cross-scale fusion.
• Introduced an interlaced concatenation method for effective combination of multi-scale cost volumes.
• Developed an adaptive cost-volume-filtering (ACVF) loss function for direct cost volume supervision using ground truth and teacher network probability distributions.

Main Results:

• The proposed MSFFNet demonstrates superior efficiency compared to previous stereo matching methods.
• Achieved 1.01 EPE, 42 ms runtime, 2.92M parameters, and 97.96G FLOPs on the Scene Flow test set.
• Outperformed PSMNet by being 89% faster, 7% more accurate, and using 45% fewer parameters.

Conclusions:

• The MSFFNet offers a computationally efficient and accurate solution for stereo matching, suitable for mobile applications.
• The ACVF loss function effectively regularizes cost volumes, improving disparity estimation in challenging areas.
• The proposed method represents a significant advancement in real-time, high-performance stereo matching.