Resources

Bibliography

Biologically motivated object recognition (MATLAB)

This is an implementation of the 'bypass' version of the standard model for object recognition. The implementation follows "Serre, T. and Wolf, L. and Poggio, T.", "A new biologically motivated framework for robust object recognition",CVPR,2005. Please visit CBCL for several publications describing this model.This version allows you to arbitrarily replace each layer with several faster approximations.

There are other implementations (C,C++,matlab) available from the following locations

Oscar Serra's matlab implementation and openGL GPU accelerated version based on Mutch's improvements
Serre et al's CVPR 2005 version
Minjoun Kouh's implementation of the Full model (The CVPR version implements the bypass route)
Jim Mutch's implementation improves upon the bypass route through feature selection and sparsification
Maximilian Reisenhuber's version is the precursor to all the above!

Description

You can find the description and trade-off of these approximations in the draft report found here

Code

The following lists the important files in the implementation.

Baseline code
Approximation code
Caltech 101 scripts
OSU SVM Classifier .This can also be directly downloaded from Sourceforge.
README.txt provides a detailed file list and their description

Data

The caltech 101 database can be found here

Installing and running the code

Download the {code,approx,caltech} tar balls and unzip them into your directory
gunzip code.tar.gz && tar -C code -xvf code.tar
gunzip approx.tar.gz && tar -C approx -xvf approx.tar
gunzip caltech.tar.gz && tar -C caltech -xvf caltech.tar
gunzip third_party.tar.gz && tar -C third_party -xvf third_party.tar
Startup matlab and add the paths
addpath code; addpath approx; addpath caltech; addpath(genpath('third_party'));
Running a binary classifier
Set the path of the training and testing data in demo_baseline.m
Running the model over caltech 101 database
Extract features over all images using script101.m script. This script is parallelized and can be run from multiple nodes simultaneously. It uses lock files to run concurrently on the whole database.
Combine the results using script101_combine_matrices
Use script101_perf.m to evaluate the performance

Biologically motivated object recognition (OPENCV)

The "bypass route" of the model has also been written in C++ using OPENCV for primitive image operations. The outputs of the C++ version are nearly identical (correlation > 0.99) with the matlab version. Additionally, the c-version is multi-threaded and computes several filter responses in parallel making it much faster than the matlab version on multi-core machines.

Code

The source code for the OPENCV version can be downloaded from here. Supporting files to convert models and filters between matlab code and C can be found here.

Sample data

The C version is currently trained to recognize four object categories (faces,mugs,books,pens). The training and testing data were collected by Johnathan Harel at Caltech and can be downloaded from here

Installing and running the code

In order to run the C++ version, you need to install opencv first. You can obtain the source distribution of opencv from sourceforge. After you install opencv you can use build_linux.sh or build_pc.sh to build the binaries(needs cygwin on PCs). Run the program without any parameters to see the proper syntax.

Training for other object categories

Currently, the C++ version performs only recognition. The training weights and features are obtained offline, using the matlab code (see above). Matlab scripts to export the S1 and S2 fitlers are provided here. You need to download the matlab version of the code and also the STPR toolbox to train the model. Use the following procedure to use the code with other object categories

Download the matlab version of the model (see above)
Use either the universal dictionary or object specific dictionary to generate features for the entire training set
Use STPRtoolbox multi-class svm (oaasvm) to train a simple classifier to distinguish between the object categories
Optionally, you can use feature selection (highly recommended) to reduce the run-time (linear in the number of features). We use L0 norm SVM to perform our feature selection (code).
Use the conversion scripts filter_m2c.m,patches_m2c.m and model_m2c.m to export the S1 filters, S2 filters and the SVM model to text files to be loaded by the C++ program.

Fingerprint Toolbox

I'm currently writing a bunch of routines for fingerprint verification in matlab. From the many mails I've got so far, it looks like anybody who wants to work on fingerprint recognition has to start from the scratch. The toolbox is my effort to provide people with implementation of popular algorithms found in literature. I will be posting the routines here as I go along. You'll need Matlab Image Processing Toolbox in order to run these programs

Some test data for working with the toolbox can be obtained from the following sources

CUBS Fingerprint Feature Extraction Tool

This tool provides a graphical user interface for minutiae feature extraction and visualization. It also allows the user to manually identify new minutiae or remove spurious ones. This tools was used by us to evaluate the efficiency of different feature extractors.The windows setup program and sample data can be downloaded from here

CUBS Fingerprint Recognition Tool

This tool provides a graphical user interface for fingerprint recognition. The tool enables visualization of matched minutiae. It also allows the user to identify new matched pairs or remove false correspondences. The windows setup program and sample data can be downloaded from here
. NOTE: The feature extraction code was contributed by Chaohang Wu (cwu3@buffalo.edu). Pl address request for feature extraction code to him. The matching and enhancement code is avaialable here

Speaker Verification

This is the source code and report for my DSP class project at UB. The source code implements Furui's cepstral features based speaker verification. The matlab source code can be downloaded from here

Speaker ID

This is the source code and report for my machine learning class project at MIT. The source code implements MFCC and VQ/GMM based speaker identification system. Also provides an option to learn optimal number of cluster centers based on Bayesian Information Criterion. The matlab source code can be downloaded from here. The zip file also contains the project report that you may find helpful

Oclet: Character recognition using wavelets and neural networks

This was my undergraduate project and implements a simple wavelet feature based character recognition. You can find the windows binaries and description here and here

Kohenen Self Organizing Map

C source code for a configurable KSOM. Source code here

Bits stream handler

A group of functions to handle streaming data (e.g. encoded audio or video stream). Source code can be found here