User:Neilh

Some relevant pages

• Neil - Research log

• Neil - TMS ideas

• Clustering people based on face, hair, and clothing

• To-do for feature detector send-off

• Neil's feature detector tech report

• Information on YGP and Love.com data sets

• VisionTiki test bed

Seminar notes

• Nonlinear dimensionality reduction seminar - Jose Costa

Research interests

• Fully-unsupervised (or semi-supervised, with only a few marked samples) visual learning
  • Multibody factorization
• Attention
• Log-polar representations
• Folksonomy/tagging: "Collaborative categorization using freely chosen keywords"
• Neural implementations
• Data mining from huge image/video data sets

Brainstorming

• Flickr has a new feature for finding tag clusters. For example, here are the different clusters for school, person and car. It may be worth looking into this more, as a way to distinguish different contexts for a particular word.
  • Another neat Flickr feature is notes, which allow you to attach descriptors to image regions.
• Is there any way to generalize the person discrimination problem to that of improved within-class discrimination? Perhaps if it learned which clustered keypoints were most crucial for within-class classification?
• Automated organization and information extraction (data mining?) of absurd amounts of image/video data, as one might see in a CyborgLog or a large flickr gallery.
  • Sousveillance
• Could use amazon.com's Mechanical Turk to get cheap annotation of plenty of images
  • One of Neil's friends from college actually works on this project
• SVM-style representations for prediction markets
• How to compute eigenvectors with neurons? (Already models of PCA with neurons.)
• Comparing deformable shapes by measuring feature distances using forest-fire methods
• Use this for background samples: http://www.vision.caltech.edu/Image_Datasets/background
• Psychophysics of motion-based supersampling
• Syntactic/structural pattern recognition
  • http://www.cs.unb.ca/tech-reports/files/tr04-165.pdf
• Compare variance of image properties for different flickr tags
• LDA-SIFT
• Attention required for boolean combinations of natural stimuli? For example, can animals and vehicles be detected simultaneously?
• Geodesic Intensity Histogram -> Geodesic Texture/Color Histogram
• Spectral mixture models; eigenmixtures
• Useful code for textures/segmentation: http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/
• Apply string kernels to images

Misc.

• Personal prediction market
• Website for wiki/discussion/sharing of research papers, keeping track of Bibtex files
• Amazon's Mechanical Turk for podcast transcriptions
• Set up a versioning control system like SVN on one of the vision machines. Be sure it isn't on a network drive, as they don't have the latency guarantees needed by a versioning system.
• TA Spring: nuSketch battlespace, forbus et al spatial reasoning, Michael Buro, def/air only