Adaptive Aggregate Stereo Matching Network with Depth Map Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Reconstruction

Background:

• Traditional stereo matching often involves direct depth reconstruction, which can limit accuracy.
• Existing methods may struggle with computational efficiency and precise disparity estimation.

Purpose of the Study:

• To propose a novel stereo matching network, ASR-Net, for improved depth map accuracy and reconstruction speed.
• To develop a coarse-to-fine approach integrating multi-level residual optimization and depth map super-resolution.

Main Methods:

• Utilized a U-Net feature extractor for multi-scale feature extraction.
• Reconstructed global disparity at low resolution and regressed residual disparity at higher resolutions.
• Incorporated deformable convolution and group-wise cost volume for adaptive cost aggregation and employed ABPN for refinement.

Main Results:

• Achieved excellent speed and accuracy on Scene Flow, KITTI 2015, and KITTI 2012 datasets.
• On KITTI 2015, attained a 2.86% three-pixel error rate.
• Demonstrated significant speed improvements, being 6x faster than GC-net and 2x faster than GWC-net.

Conclusions:

• ASR-Net effectively refines depth maps through a coarse-to-fine strategy, avoiding direct reconstruction issues.
• The network offers a superior balance of speed and accuracy compared to state-of-the-art stereo matching techniques.