HMAX Package for CNS

Description

This is a CNS package that can be used to instantiate a broad class of feedforward object recognition models. The package includes demo scripts to build and run the specific models described in these papers:

• Jim Mutch and David G. Lowe. Object class recognition and localization using sparse features with limited receptive fields. International Journal of Computer Vision (IJCV), 80(1), pp. 45-57, October 2008. [pdf]
• Jim Mutch and David G. Lowe. Multiclass Object Recognition with Sparse, Localized Features. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11-18, New York, June 2006. [pdf]
• T. Serre, A. Oliva and T. Poggio. A feedforward architecture accounts for rapid categorization. Proceedings of the National Academy of Science, 104(15), pp. 6424-6429, April 2007. [pdf]

Installation

See the instructions for installing packages in the CNS manual. (CNS itself must be installed first.)

Once the package has been installed, make sure it's working by running the following demo script (described in more detail below):

>> hmax_cvpr06_run_simple

The script should output the following lines. The CNS manual contains a troubleshooting section if you encounter problems.

sampled 2048 "s2" features from "image_0010.jpg" computed feature vector for "image_0002.jpg" (??? sec)

Introduction

Each of the feedforward recognition models described in the above papers computes a hierarchy of increasingly complex features that are increasingly invariant to position and scale. This is done in interleaved stages that perform either tuning or pooling. Tuning stages are named "S1", "S2", etc. and pooling stages are named "C1", "C2", etc. Each stage is computed at multiple resolutions (scales), i.e. each stage is a scale pyramid. (This is best illustrated in figure 2 of this paper.)

In this hmax package, each of these stages is represented as a CNS group, composed of multiple CNS layers, one layer per scale. This allows all the scales of a group to share common parameters and (if applicable) a common feature dictionary. All the units or cells in a group are of the same type (dot product, max pooling, etc.)

Under the directory hmax/demo there are several demo scripts.

hmax_cvpr06_run_simple
hmax_pnas07_run_simple

hmax_cvpr06_run_cal101
hmax_cvpr06_run_uiuc

hmax_cvpr06_params_base
hmax_cvpr06_params_full
hmax_pnas07_params

The hmax package itself consists of the file hmax.m plus many pairs of files named hmax_type.m and hmax_type.h, each of which defines a different cell type. Some of the ".m" files also contain useful methods, many of which you will have seen used in the above demo scripts.

• hmax.Model - converts a compact HMAX parameter set into a full CNS model structure which you can then instantiate and run on the GPU.
• hmax_input.Load - loads a new image into an input layer of an instantiated model.
• hmax_s.SampleFeatures - adds new features to a dictionary by sampling.
• etc.

Additional Documentation

Documentation for this hmax package currently consists of:

• this page
• the demo scripts mentioned above
• MATLAB function and method help
• comments in source files

Revision History