Datasets
Kyoto natural images
A set of calibrated color natural images, including three classes of cone photoreceptor responses (i.e., long (L), medium (M), and short (S) cone types) with their nonlinear transform.Go to the kyoto images page.
Pittsburgh Natural Sounds
A set of sound files recorded around the Pittsburgh region combining ambient sounds (rain, wind, streams) with acoustic transients (snapping twigs, breaking wood, rock impacts).Go to the sounds data page.
Aluminum (left) and wood. Click on images to hear samples.
Impact sounds
A set of impact sounds that were produced using four rods with the same length and diameter but made of different materials. A wooden rod, with a much shorter length but the same diameter, was used as a mallet. Several impacts were recorded (in an anechoic chamber) for each rod, and the location of the impacts as well as the impact force varied slightly from one instance to the other. The sounds were digitized using a sampling frequency of 44100 Hz.Data coming soon...
Code
Gabor-fitting routines
MATLAB functions to fit parametric 2D Gabor functions to square image patches. This implementation makes use of the (commercial) TOMLAB optimization routines for MATLAB (we found them to be more efficient at finding global optima of the optimization). These solvers (TOMLAB's base module) are similar to MATLAB's own optimization toolbox functions, so it shouldn't be too hard to convert these scripts (but we'll leave that to you).
The 2D Gabor function is a product of a 2D Gaussian bump and a 2D
sinusoid,
The fitted parameters are the center (μx, μy), orientation (θg), and scale (σx, σy) of the Gaussian; the frequency (ω), the phase (φ), and the orientation (θs) of the sinusoid; and the amplitude (h) of the final Gabor function.
The fitted parameters are the center (μx, μy), orientation (θg), and scale (σx, σy) of the Gaussian; the frequency (ω), the phase (φ), and the orientation (θs) of the sinusoid; and the amplitude (h) of the final Gabor function.
Download the tar.gz archive.
This code was developed by Mike Lewicki and Eizaburo Doi.
Density Components
MATLAB code for Maximum Likelihood estimation of Density Components, as described in Karklin and Lewicki, "A hierarchical
Bayesian model for learning non-linear statistical regularities in
non-stationary natural signals", Neural Computation, 17 (2): 397-423,
2005. [pdf]
Download the MATLAB file.
(This is just a skeleton script to illustrate the algorithm and the gradient computations. It does not include routines for data loading or pre-processing, or for visualization and analysis of the results.).