Cacher is the code snippet organizer for pro developers

Weekly Meeting

Meeting minutes

During the meeting we discussed about the overview of SfM and in our case we can utilised the similar approach to build our own pipeline in order to realise the tracking component of our AR project. An interested discussion during the meeting is that we agreed on how traditional tracking algorithm is not exactly what we are looking for. Instead, our tracking needs to deal with geometric transform up to "prespective" level. Also, majority of state-of-art on camera motion estimation rely on finding the optimal solution for correspondance matching. Our argument point, or perhaps our thought is that, for a given correspondance result we can ramdomly extract 4 lines and deduce a homography, "Is there anyway to pick up the most highest confidence score 4 lines from the correspondance pool"?

Things to do

• Take a short street view footage, and use this video to the demo programs (TLD, fern and optical-flow based) to see the overall result.
• Ingestigate in "Ranking the best correspondance". The concept behind it is to deduce a homography using those with "highest" score so the result could be improved compared to the tradidional "using all points" especially in case there are false inliers matches.
• Investigate in "Fast Keypoint, 2010", especially for the training part.

Tutorial resources

• part 3: should i read it? the naive bayes method
• v1: semi supervised naive bayes

Useful Resources

• An Effective Rigidity Constraint for Improving RANSAC in Homography Estimation

  Research paper on adding transformation constraint in RANSAC homography estimiation. This solve my problem in determining the good homography.

• Speeding-up homography estimation in mobile devices

  Research paper on homography estimiation tracking, haven't read it yet!

• Fast Keypoint Recignition in 10 lines of code

  Research paper on the binary festure descriptor, it is used in "Fast KeyPoint, 2010".

• Visual Object Tracking via random ferns based classification

  Research paper on object tracking using fern descriptor, used in "Fast Keypoiny, 2010".

• A simple explanation of Naive Bayes Classification

  General discussion, intuitive explanation on Naive Bayes Classification.

• http://www.saedsayad.com/decision_tree.htm

  General discussion, example of building up decision tree.

• Structure from Motion

  A long lecture note on structure on motion.

Programming Library Resources

• Theia Lirary

  A C++ library to provide SfM implementation, haven't personally try it yet but sounds interesting.

• Matlab Computer Vision tool

  A set of useful computer vision functions by Peter Kovesi.

• Dlib

  Lightweight library that provide machine learning and computer vision functions. I am currently using this as part of C++ implementation for master project.

Previous record

20160906