CS129 / Project 6 / Automated Panorama Stitching

interested points demonstration

Overview

I implementted the basic version plus the rotation invariance. Here are the details:

  1. Non-maximal suppression: I use 0.9*neighbor as threadhold and take top 500
  2. Feature descriptor extraction: I implemented a rotation invariance version of this part. I added a function patch_gradient() to calculate the angle to rotate and also a parameter 'rotate' 'to get_features(). Rotate =1 means rotating the patch to the direction of its dominant graidient. In the result table, I compared the oriented ones with naive ones.In most cases the results are similar, except for source007, where the oriented one looks better than the naive one.
  3. Feature matching: I have tried 0.6,0.7,0.8,0.9 for the Lowe's ratio test. I also take into consideration of the graph in the lecture slide. My end result is based on 0.7 ratio as threshhold.
  4. I used the suggested 1000 iterations but adjusted the error slightly(to squared error of 0.5).

Source007 oriented results and naive one(left: naive one, right:oriented feature vector one)

Rest of the pictures. They are all orented ones. However, most of them are not that much different from naive ones.The last two images are taken by myself using Sony Nex 5