|
Here is my ICCV 2005 contest entry form. Here is a link to the contest description.
The contest is over, but you can view my progress log below,
or carry on to my project page, where I am continuing to work on my entry
for a course on nonlinear optimization. I have also compiled a
list of the approaches of the top 5 teams.
I see this contest as an excellent opportunity to meet and network with other young vision researchers.
I am amenable to collaboration, and would be more than happy to share/discuss the problem and my solution ideas -- please
email me if you are of like mind :).
A short Matlab script to generate Google map links about specified latitude/longitude coords.
Progress:
Oct-03-2005. After programming steadily for about 72 hours, the contest deadline
passed by with my entry below the continuation cutoff. I simply didn't have enough time to get my bundle adjustment
code running. Good luck to the remaining five teams. As for me, I intend on finishing my entry for my non-linear optimization
course project, so the hard work has certainly not been for naught. How my system worked:
- Initialize a graph where nodes represent images, weighted edges matches between images
- Find keypoints, extract SIFT features
- Build a kd-tree containing all points
- For each unannotated image:
- Find best match amongst all other images (max. no. nearest neighbour matches)
- Match keypoints between image and its best match
- Cull non-discriminative matches
- Estimate fundamental matrix with MAPSAC
- Connect images with an edge, weighted by inverse of number of inliers consistent with F
(add no edge if there are no inliers)
- Repeat on next-best match, until an annotated match is reached (or match limit)
- For each unannotated image, find the shortest path to an annotated one and equate their coordinates
(if no path exists, assign the best match according to
max. no. of keypoint matches alone)
- (Not Implemented) Structure and motion reconstruction by IRLS (Levenberg-Marquardt)
on camera parameters of connected subgraphs
See you at ICCV!
Oct-01-2005. Go-time! Programming frenetically.
Aug-25-2005. Making progress in 3D, but I still have a lot to implement. I am registered for the conference,
and extremely excited to go -- this will be my first ICCV.
Aug-16-2005. Perfect nearest neighbor matching on Test3, though method may be fragile on harder datasets.
3D difficult, due to wide baselines. Time scarce due to journal paper and preparations for Burning Man festival.
Aug-10-2005. Abandoned my old approach. I'm still stuck in flatland, but my Test2 results have been improved,
and the new technique works on Test3. Time has been scarce of late, because of paper crunches and the demands
of less academic summertime persuits.
Jul-25-2005. Baseline system complete, but fine-scale accuracy still low. Will robustify with
local feature matching (lots of implementation to be done).
Jun-20-2005.
I can split (autonomously, of course) Test2 & Test 3 into three disjoint subsets
corresponding to the three scenes that comprise this dataset. I anticipate my approach will succeed on Test3 with
little or no modification.
Back to the main page
|