Feature Detection and Matching

AER1515 - Perception for Robotics - UTIAS

RANSAC feature matching with horizontal contrainst
  • This project implements feature point detection and its matching between stereo pair images from KITTI dataset.
  • For a given input RGB image from left camera, the features which are described to be an image region that is salient, local, repeatable, compact and efficient, are identified and studied by visual inspection for unreliability on matching.
  • As the features are detected in its corresponding right camera image as well, we match the feature with a brute force matcher.
  • For each match pair identified, the disparity between the x values are calculated, and they are converted into depth values using the predefined functions and calibration details provided.
  • The calculated depth values are then documented as a text file and attached for evaluation on test images.
  • A RANSAC based outlier rejection method has been implemented on the match pairs, and promising results have been noted.
  • The method has been fine tuned on train images by evaluating the results for corresponding changes to its confidence and projection threshold values.

Code is available here

  • This is a python notebook .ipynb file created on Colab, then exported to a .py file.
  • The dataset is accessed by mounting the google drive. Please change the path of dataset accordingly in line 259. Also change file paths for left, right images and to access calibration files in line 262 - 265, and 288.
Mahesh Sudhakar
Mahesh Sudhakar
Computer Vision Research Engineer

Computer Vision | Robotics | Machine Learning