A Joint 2D-3D Complementary Network for Stereo Matching.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Stereo matching is crucial for 3D reconstruction in computer vision.
• Existing neural network stereo methods struggle to balance speed and accuracy due to cost aggregation complexity.

Purpose of the Study:

• To develop a stereo matching method that achieves both high accuracy and reduced running time.
• To overcome the speed-accuracy trade-off in current neural network-based stereo vision.

Main Methods:

• Integration of a fast 2D encoder-decoder network for initial disparity map generation.
• Utilizing a generated disparity range to guide a 3D aggregation network for enhanced accuracy.
• Employing a stacked hourglass structure for coarse-to-fine disparity refinement.

Main Results:

• The proposed method generates accurate disparity maps significantly faster than existing 3D stereo networks.
• Achieved a running time of 80 ms on a modern GPU for accurate results on the KITTI dataset.
• Demonstrated substantial accuracy improvements over other 2D stereo networks like AANet, DeepPruner, and FADNet.

Conclusions:

• The hybrid 2D-3D network effectively balances speed and accuracy in stereo matching.
• This approach offers a computationally efficient and highly accurate solution for computer vision tasks requiring depth perception.