This post is about the pictures I received from Greenland.  I put them through the image normalization algorithm (That uses homography to corerct things) and the results were a little disappointing :(.  There are two aspects of the transformed images that I want to mention: 1. the accuracy of the id detection and 2. the accuracy of the homography normalization.

With respect to the id detection; that is, classifying the information contained in the color chessboard; the algorithm had 100% accuracy with the pictures that I received.  The classification for all the pictures took a while (30 mins for 68 pictures, 30s/picture).  I was very happy with this behavior and think I can , for the time being, stop touching the logic for the id extraction.

The accuracy of the homography was not so smooth as the id detection.  The sections of the picture that are close to the chessboard marker normalize beautifully.  I can barely notice the difference when I go from one picture to the other.  But as you move away from the marker the correspondence from picture to picture is not that good.  It’s so bad that in some cases the same point moves several hundred pixels when one changes image.

One of the problems I detected with the current process is that the marker is in one of the corners of the image.  This means that the opposite corner to the marker will have little correspondence between images.  To make my point I have created a little clip.  While watching it compare the behavior of the part that is close to the marker and the part that is far from the marker.

To solve this issue I think it is best to put the marker in the middle of the plot.  I know this sounds counter intuitive because it would be blocking part of the plot.  But if you consider the size of the marker, there is not much information that is lost.  I’m going to try this out when I go to Greenland in half a month.