Greg's personal wiki

Current Logbooks

Olfaction           Pollen

ICCV 2009           Bird Taxonomies

Software Docs          

Old Logbooks

Spring 2008 Olfaction Notebook Pollen Notebook

Fall 2007

Summer 2007

Spring 2007

Winter 2006-2007

Summer 2006

Current Papers Topics

NIPS 2007 1. Learning And Using Taxonomies For Visual Category Recognition placeholder 2. Using Phylogenetic Techniques to Learn Taxonomies of Visual Categories 3. Attentional Cascade for Rapidly Identifying Of Hundreds of Object Categories ie. The Interesting-ness Detector At Work conference web page submit 6/8/07

Future Conference 4. Exploiting Taxonomies of Visual Categories Using Hierarchical DAG-SVM 2007 Conference Summary

Milestones

ROC Curve for three different interest classifiers

Taxonomy learned from confusion matrix

Performance comparison

Most of the results below use the Spatial Pyramid Matching algorithm of Lazebnik, Schmid and Ponce (CVPR06).

Jan 2006:
  Implemented fast classifier for interesting-ness

Dec 2006:
  Wrote hierarchical top-down category classifier
  (20,000 lines of MATLAB code so far)

Nov 2006:
  Writing Learning And Using Taxonomies
  For Visual Category Recognition

Oct 2006:
  Phylogenetic algorithms successfully learn
  taxonomies using only the confusion between categories

Sep 2006:
  $67.2\pm 1.0$  on Caltech 101  (N_train=30)
  $~34.1\pm 0.1$  on Caltech 256 (N_train=30)

Aug 2006:
  84% true positives with 0.5% false positives on cat
  detection where all cats are at the same scale (N_train=900)

Jul 2006:
  $63.0\pm 0.6\%$   on Caltech 101 (N_train=25)
  $80.1\pm 0.9\%$   on Scene Database (N_train=100)

Jun 2006:
  Spatial Pyramid Matching re-implementation

Previous Milestones

Research Projects

The first 64 of the Caltech-256 categories

One of over 30,000 images in the Caltech-256

Sift feature extraction

Segmentation experiments

• Caltech-256 was collected mostly by me and Alex. Questions and comments can be addressed to me.

• Re-implemented the Spatial Pyramid Matching algorithm of Lazebnik, Schmid and Ponce (CVPR06). These matching kernels are being used in several different projects around the lab.

• Top-down category recognition based on a a simplified taxonomy (ICCV2007)

• Generating taxonomies from a confusion matrix (ICCV2007). Do this the same way that geneticists generate phylogenetics trees. Use something like Lin's DAGSVM to speed up classification.

• Scanning for interesting things you've never seen before (NIPS2007) and making taxonomy-based guesses as to what they are. Then ask a human for help (like child would). The trick is quantifying what we mean by interesting (as opposed to clutter). Trees may help.

• Scanning images for cats and dogs (and everything else)

• Eigenfeatures for recognizing new categories with fewer training examples, based on what we know about our existing space of category-specific features. Another use would be to find linear combinartions of n different classifiers that are maximally tuned from discriminating each category. Every algorithm has things that it gets confused about. Eigenfeatures would try to break the degeneracies in a principled way.

• Finding Gravitational Lenses using vision algorithms. Hien Nguyen at JPL studies gravitational lenses. Human identification is time-consuming and slow, but they are very difficult to find using conventional computer methods. This may make it a perfect task for us. If they are found, follup observations can be made for confirmation.

• Investigating Improvements to Pyramid Matching such as clutter removal, flexible bins boundaries and binning in both pixel and feature space.

• Other Techniques for Image Matching

• Iterative segmentation based on maximizing information. Ultimately it seems like you need to separate the object from the clutter in order to get better performance. Can we do this iteratively by minimizing some cost function?

• Radially Binned Fourier Transforms: borrow a simple feature set from astronomy and ask, how well does it work compared to SIFT and gaussian blur functions? Describe how telescope scan patterns and computer-based features are created to address the same problems. What other methods can be borrowed from astrophysics?

Related Projects

• Active learning is one of the projects Alex is working on. These are some notes on how to generate Pyramid Matching Scores that are geared towards that project.

Cat detection experiments

• Ryan Gomes is developing new techniques for detection and incremental learning, especially for the case of unsupervised datasets. You can find out more on his page. He's using some of the Caltech-256 Pyramid Matching matrices for unsupervised classification based on KPCA [1] [2]. In the real world you can't label hundreds of thousands of images by hand. Even if you could, you wouldn't have infinite amounts of memory to remember them all. These restrictions are becoming more and more important as we go from toy experiments to dealing with hundreds or thousands of categories.

• Francois Fleuret  works on top-down detection of faces and more complex objects. We compared a couple methods for detecting cats. Francois got a really huge database of cats from ratemykitten.com. Now.. what to do with it?

• The Fade Cascade is an attempt to allow decision cascade algorithms (e.g. Viola and Jones face detection) to learn with fewer training examples. Descision cascades are an important tool for fast top-down detection, so any improvements would help. This is a collaboration with Jeff Edlund; we hope to write up a paper sometime soon.

• Here's an interesting psychometric experiment by someone in the lab, on how children acquire visual taxonomies.

• I worked on this Pyramid Matching project with Will Coulter in EE148.

Presentations

Jan 2007: Here are a few slides for Pietro

May 2006: In CS186 Will Coulter and I gave this final talk PDF PPT

Mar 2006: Introductory talk for CS186 presenting the Pyramid Matching algorithm of Lazebnik et. al. PDFPPT

Jan 2006: Jeff Edlund and I proposed a Fade Cascade for our project in CS 129b.

Old Material

---

Click here to email me

---