Project 2: Local Feature Matching

Effects of design decisions

After implementing get_features and match_features my initial accuracy on the Notre Dame pair was around 50% with the cheat points. Initally I was only normalizing each feature to unit length. Thresholding the feature to range .2 and then normalizing again pushed the accuracy to 68%. For each feature I if there was a negative value present in the vector I would set it to 0. This however was a lossy way to reduce points with bad votes. Instead I replaced negative values in the cells with 0. This change increased the accuracy to 71%. After implementing the harris dectector the best accuracy for the Notre Dame pair was 83%.

Results

Notre Dame: 83% Accuracy

this is image #1 this is image #2

Mount Rushmore: 95% Accuracy

Episcopal Gaudi: 7% Accuracy

I was able to acheive decent results on the Notre Dame and Mount Rushmore image pairs. Episcopal Gaudi seems to require extra work to achieve decent results.