Stereo Matching Algorithms in MATLAB and Python

Optimized (very fast) stereo matching algorithms in MATLAB and Python. It includes implementations of Block Matching, Dynamic Programming, Semi-Global Matching, Semi-Global Block Matching and Belief Propagation.

Features

• Stereo matching algorithms

  • Block Matching (BM)
  • Dynamic Programming (DP)
  • Semi-Global Matching (SGM)
  • Semi-Global Block Matching (SGBM)
  • Belief Propagation (BP)

• Multiple different versions of the algorithms

• All algorithms are implemented in both MATLAB and Python

• The algorithms are optimized for performance using matrix operations and other techniques

Algorithms

Filename	Description	Implementations
stereoBM_SAD	Block Matching using Sum of Absolute Differences	MATLAB / Python
stereoBM_SAD2	Block Matching using Sum of Absolute Differences (different approach)	MATLAB / Python
stereoBM_Grad	Block Matching using Image Gradients	MATLAB / Python
stereoBM_NCC	Block Matching using Normalized Cross-Correlation	MATLAB / Python
stereoBM_Rank	Block Matching using Rank Transformation	MATLAB / Python
stereoBM_Census	Block Matching using Census Transformation and Hamming Distance	MATLAB / Python
stereoBM_ASW	Block Matching using Adaptive Window (Adaptive Support Weights)	MATLAB / Python
stereoDP_LR	Dynamic Programming with Left–Right Axes DSI	MATLAB / Python
stereoDP_LD	Dynamic Programming with Left–Disparity Axes DSI	MATLAB / Python
stereoSGM	Semi-Global Matching	MATLAB / Python
stereoSGBM	Semi-Global Block Matching	MATLAB / Python
stereoBP_Synch	Belief Propagation with Synchronous Message Update Schedule	MATLAB / Python
stereoBP_Synch2	Belief Propagation with Synchronous Message Update Schedule (different approach)	MATLAB / Python
stereoBP_Accel	Belief Propagation with Accelerated Message Update Schedule	MATLAB / Python

Installation

Download the project as ZIP file, unzip it, and run the scripts.

Python Requirements

• NumPy
• Matplotlib
• OpenCV (opencv-python)

Usage

A stereo matching algorithm works with stereo image pairs to produce disparity maps. This project contains MATLAB and Python scripts, each implementing a stereo matching algorithm. The files left.png and right.png contain the stereo image pair used as input. To use a different stereo pair, replace these two images with your own. In this case, you must also adjust the disparity levels parameter in the script you are running. You may optionally modify other parameters as needed. If the input images contain little or no noise, it is recommended not to use the Gaussian filter.

• The results between MATLAB and Python implementation are similar.
• The different approaches produce same results.

Performances

The following running times were measured on a Windows PC with a CPU AMD A10-7850K and 8GB of RAM.

Filename	MATLAB (sec)	Python (sec)	Notes
stereoBM_SAD	0.10	0.08	image display disabled
stereoBM_SAD2	0.48	0.81	image display disabled
stereoBM_Grad	0.17	0.19	image display disabled
stereoBM_NCC	1.52	1.84	image display disabled
stereoBM_Rank	0.77	3.70	image display disabled
stereoBM_Census	2.94	7.25	image display disabled
stereoBM_ASW	20.58	46.11	image display disabled
stereoDP_LR	0.32	5.60	image display disabled
stereoDP_LD	0.14	0.28	image display disabled
stereoSGM	0.81	2.45	image display disabled
stereoSGBM	0.84	2.51	image display disabled
stereoBP_Synch	11.16	21.00	image display disabled, 20 iterations
stereoBP_Synch2	4.83	17.57	image display disabled, 20 iterations
stereoBP_Accel	9.11	15.18	image display disabled, 20 iterations

Results

Below are the disparity maps produced from the Tsukuba stereo pair.

Block Matching (SAD)

Block Matching (Gradient)

Block Matching (NCC)

Block Matching (Rank)

Block Matching (Census)

Block Matching (Adaptive)

Dynamic Programming (Left-Right)

Dynamic Programming (Left-Disparity)

Semi-Global Matching

Semi-Global Block Matching

Belief Propagation (Synchronous)

Belief Propagation (Accelerated)

Links

• Project repository: https://github.com/aposb/stereo-matching-algorithms

License

This project is licensed under the MIT License. See the LICENSE file for details.