University Honours Project

I’ve added my university honours project to this website. It is called Modern to Historical Image Feature Matching. The project can be briefly described as an attempt at understanding why modern computer vision techniques aren’t successful at matching historical photos taken on film with modern digital photos. The project includes a paper (linked above), source code and a data set. The project was written in C++, using the OpenCV computer vision libraryAll the appropriate links can be found in the programming section of hd konulu porno sikisxxx. The following is the abstract of the project paper:

Matching images with each other using their unique features has many applications. While there currently exists many robust image
feature matching systems, we propose that they are currently only accurate when matching modern photos with each other. Historical
photos, traditionally taken on film, when converted to digital format are not likely to successfully match with modern equivalents. Photos of famous landmarks from various dates were gathered to verify this hypothesis. Images were matched using four standard feature matching systems. These systems matched either image corners or image blobs with each other. Match successes were recorded for each image comparison, and analyzed in tables. The results showed that the feature systems accurately matched modern photos with each other. However, when matching historical photos to modern photos, results were poor. A novel approach to feature matching using image regions was proposed. The expectation was that using regions instead of corners or blobs would result in higher accuracy when matching. Initial results were mixed, and no concrete solution was found.

Posted in Programming | Tagged , , , | 2 Comments | 2,481 views


  1. andrew


    I found your paper very interesting. Since the existing features seem to be unsuited to matching in historical images what would you think might work?

    Have you tried Hessian-Affine points (instead of circles you have elliptical patches)?

    Are the pictures listed in the paper the only ones you tested on? I’m guessing if it didn’t work so well on these, where the photos are taken from a single view, it would work even worse for different views.

    Good luck with your future work!

    1. Robbie Wolfe Post author

      My supervisor thought MSER was the right approach, but this project proved unsuccessful. It’s still an open problem.

      I tested on 10 landmarks, which each had 10 images that I found on the internet. I had two challenges, I had to find images from similar angles, even though they were obviously taken by different people. I also needed images with date information, otherwise it wasn’t useful for this project. Some images weren’t exactly the same angle, which could have definitely influenced the results. However, they all had similar shapes/regions, so the standard feature detectors should have worked better than they did.

      Thank you for reading my paper, I’m glad you found it interesting.

Comments are closed.