Computer Vision: State Of the Art

1. Key point detector
SIFT
ASIFT
SURF

David Lowe reported that SURF is more rebust than SIFT in affine transformation, but SIFT is more robust to 3D rotation.

2. Background subtraction
No state of the art. Though I prefer Mixture of Gaussian personally.

See this
http://www.andrewsenior.com/technical/surveillanceclass/02_MovingObjectDetection_part1.pdf

3. Salient motion detection

see this
http://www.andrewsenior.com/technical/surveillanceclass/03_AdvancedMovingObjectDetection.PDF

4. Stereo

see this
http://vision.middlebury.edu/stereo/eval/

5. Robust estimation

RANSAC -Good if there are many inliers. If not, it needs lost iterations to find a good match
Hough Transform -Use this when RANSAC runs too slow
semi-local constraints -Check the consistency with nearby keypoints

The difference of RANSAC and Hough Transform lies in computation time. Try both to see which one runs faster.

Todo: expand the list, provide all implementations
Reminder: don't spend too much time on paper crawling, do some real stuff

6. Source of training data
ImageNet: A Stanford project to label image by volunteers
http://www.image-net.org/
 Caltech 256
http://www.vision.caltech.edu/Image_Datasets/Caltech256/

Surveillance data: (have ground truth??)
http://pets2006.net/
http://www.elec.qmul.ac.uk/staffinfo/andrea/avss2007_d.html
http://www.nada.kth.se/~hedvig/data.html
http://scienceandresearch.homeoffice.gov.uk/hosdb/cctv-imaging-technology/i-lids/pricing.html

7. Human Attention
A lab of USC doing this. By Prof. Laurent Itti and Prof. Christof Koch
http://ilab.usc.edu/bu/