Peter's SVR Notebook

From Vision Wiki
Jump to navigation Jump to search

ICCV 2009 Paper: Left to do

  • Verify that re-ranking works for some datapoints.
  • Look at how LSH speed varies with database size.
  • How do the no. of model-matches (to correct model) drop off with database size?
  • 500K vocabulary for houses.
  • Fair plot for Philbin.
  • Why does Philbin break down.
  • Understand why optimization for LSH didn't work, and why it works now. Need for discussion.

Core Message

  • Which method is good to use of SIFT?
  • Which one scales the best?
  • Which one is fastest?
  • Which one is most accurate?

Speed Issues

  • For Philbin's approach, is the computation really logarithmic? How many matches do we get for different database sizes?

Memory Issues

LSH

  • What are the inherent problems in the LSH algorithm for SIFT descriptors?
    • Cannot have R too small, as we will miss good NNs.
    • Cannot have R too large, as we will get too many bad NNs.
    • In practice, this will lead to linear scaling in computation time for large databases.
  • Try using fixed L-prime and see how accuracy degrades (more realistic setting).
  • Sanity-check matches to see that inliers are really inliers.
  • Why didn't original tuning work?
    • It is important to pick a dataset with realistic query points -- i.e. not spurious points far from everything, but points that acutally survived RANSAC.
    • Show how 3 sets of points give vastly different parameter settings. (pts surviving ransac, pts NOT surviving ransac, and no such consideration at all).
    • Also need low prob -- 90% is ridicolous.
    • How about R.

Philbin