Stereo Example for POLAR Traverse Dataset

This repository contains a simple example of OpenCV's SGBM stereo algorithm run on the POLAR Traverse Dataset.

Data

The images used for this simple example are from the following subset of the dataset:

• View 1 (forward-facing, light at 20 degrees)
• Traverse 3 (cameras at 1.35 m height with 35 degree pitch)

5 total image pairs are used with the following additional parameters:

• 1 m along test bed (locs 46 and 47), 25 ms exposure
• 9 m along test bed (locs 62 and 63), 5 ms exposure
• 9 m along test bed (locs 62 and 63), 25 ms exposure
• 9 m along test bed (locs 62 and 63), 75 ms exposure
• 9 m along test bed (locs 62 and 63), 300 ms exposure

The images do not have the same names as they do in the POLAR Traverse Dataset (this example was created before the naming conventions for the published dataset were defined).

Example

The gen_disparity.py script in the scripts subdirectory can be used to produce the results. This script calls the stereo matcher class defined in stereo.py to perform image rectification and matching using OpenCV, followed by computation of disparities and generation of point clouds.

The input data is contained in the data subdirectory. Results will appear in the results subdirectory.

The stereo algorithm parameters and the camera calibration information are found in the cfg and calib subdirectories, respectively.

Depends

This code was run using Python 3.8.10 on Ubuntu 20.04 with the following additional packages:

• OpenCV (stereo) >= 4.2.0
• Open3D (point cloud processing) >= 0.17.0
• numpy >= 1.24.3
• matplotlib >= 3.7.1
• yaml >= 5.3.1