HMAX

M. Riesenhuber and T. Poggio,

"Hierarchical Models of Object Recognition in Cortex,"

Nature Neurosciece, Vol. 2, No. 11, November 1999. pp 1019-1025.

They say simple cells and complex cells and

complex cells are responsible for invariant properties.

Invariance can be implemented as a pooling mechanism

where there are view-invariant units and view-tuned units.

Hierarchical feedforward network is considered and

the network is based on MAX rather than linear summation (SUM).

MAX is proved to be more robust and invariant than SUM.

That is, Hierarchical MAX has view-invariant object recognition ability

and is supported by biological (physiological) facts.

Anyway, I say, this is a very similar concept to ISA, but a little more flexible.

But this paper doesn't say much of implementation such as

how to get the simple cells or complex cells,

and how to construct the network structure.

See [1] for a specific implementation and examples of HMAX.

In [1], interestingly, simple and complex cells are not learned from data

but designed by second derivative of Gaussians.

[1] T. Serre and M. Riesenhuber,

"Realistic Modeling of Simple and Complex Cell Tuning in the HMAX Model,

and Implications for Invariant Object Recognition in Cortex,"

AI Memo 2004-017, CBCL Memo 239, MIT, July 2004.

- H. Choi

"models of object recognition"

M. Riesenhuber and T. Poggio,

"Models of Object Recognition,"

Nature Neurosciece Supplement, Vol. 3, November 2000. pp 1199-1204

It's a review paper about object recognition models.

Here is what I got from the paper.

"the distinction between identification and categorization is mostly semantic."

There are two kinds of models for object recognition.

1. view-based model: "objects are represented as collections of view- specific features"

It is something like ICA.

2. object-centered model: there is 3-D model of the object.

It's something like Geon theory.

They took "view-based model" in this paper.

Considering the speed of processing in the brain,

feedback model cannot be the prime model for object recognition.

It should be more like feedforward processing

where invariant properties can be obtained by hierarchical structure.

Making the connections between input image to higher level units as in Fig. 3,

the different tasks (categorization and identification) can be achieved by learning.

- H. Choi

a response to my review...

Dear reviewer,

This is just to inform you that the author of the above manuscript, for which you have provided a review, has asked us to convey their thanks for your valuable comments.

It's my pleasure. :)

Actually I have reviewed journal papers since 2006, but I've never got this kind of response.
It's so kind of the author and I hope the paper gets published soon.. :)

- H. Choi

Complex Cells and Object Recognition

S. Edelman, N. Intrator and T. Poggio, "Complex Cells and Object Recognition," NIPS97.

Note that this paper was published in 1997, which is 'long time ago.'
It says complex cells-like filter has invariant recognition. And, actually, it is really simple. Apply one filter for complex cells and check the correlation of the filter output to classify.

The problem is how to implement the complex cells-like filters. And we have one answer, which is independent subspace analysis (ISA) which is kind of a generalization of independent component analysis (ICA). BTW, ICA is a filter like simple cells.
The real problem is these theories are not like math theories. So the performance really depends on the situations such as noise or background or the shape of object.

Our brain is sooooo amazing... How does it do all these complicated stuffs?

- H. Choi

Active Control..

K. L. Harman, G. K. Humphrey, and M. A. Goodale, "Active Manual Control of Object Views Facilitates Visual Recognition," Current Biology, Vol. 9, No 22. pp1315-1318. 1999.

To sum up, first, observers who have active control on the object have better recognition than others who see the same sequence passively. The response time is shorter but accuracy is the same which is a little ackward. Second, active observers concentrate on certain angles than others. That is, this paper is kind of the first paper to emphasize on the active control.
But the reason why it is efficient is still a research topic to go.

In a word, active control is important on recognition, which now we all know and agree to... :)

- H. Choi

3D Geon Classification

W. Xing, W. Liu and B. Yuan, "A Novel Integrated Scheme for Extracting Superquadric-based Geons from 3D data," in Proc. of ICSP'04, 2004.

To my knowledge, this paper is the first paper to try to implement Geon Theory, even though this is just about Geon classification. They build up a superquadric-based model for 3D object data and optimize the parameters. And then they apply SVM to the extracted features from the parameters of the model to classify the object into geon classes.

This is not about 2D image but 3D data. And this makes a model for each 3D data and extract some features from the model parameters, which might be rotation and transition invariant. This does not sound natural and not a visual recognition.

- H. Choi

Geon Theory and Its Implementation (?)

K. Casey and C. Exton, "A Java 3D Implementation of a Geon Based Visualisation Tool for UML," PPPJ 2003, Kilkenny City, Ireland, 16-18 June, 2003.

I was googling to check if there has been any kind of implementation of Geon Theory. So, when I typed in 'goen' and 'implementation,' this paper popped up to my surprise. I was thinking this theory was too hard to be implemented.

But... it turned out that this paper is just about the implementation of a visualization tool based on geon theory... not implementation of geon theory, in that the geons and their relations are equivalent to ULM diagram, which is a visual language for modeling software designs.

Anyway, so far, to my knowledge, there has been no successful implementation of Geon Theory. If you know any implementation, plz let me know. :)

- H. Choi

Three-dimensional object recognition

M. J. Tarr, P. Williams, W. G. Hayward and I. Gauthier, "Three-Dimensional Object Recognition is Viewpoint Dependent," Nature Neuroscience, Vol. 1, No 4, 1998. pp275-277.

It is a little against Biederman's RBC (recognition-by-components) which is based on geons and their relations. Biederman's geons are viewpoint independent. This paper says they are not independent.

Obviously, object recognition is a viewpoint-dependent process. So this paper makes more sense. And this paper makes a way to connect this whole things to X which means a lot to me*.
One thing more interesting is one of the assumptions of whole this geon theory, which is that recognition of individual geons are equally accurate. Actually in real pattern recognition, it is a big deal. We should handle this first, if we want this theory.

* X refers to something that I am currently working on.

- H. Choi

Visual Recognition

K. Grill-Spector and N. Kanwisher, "Visual Recognition: As soon as you know it is there, you know what it is," Psychological Science, Vol 16, No 2, 2005. pp152-160

It tests two hypotheses about processing order between detection and recognition, between basic level recognition and other level recognition. To sum up, detection (or segmentation) and recognition work at the same time and general categorization is faster than finer level identification. And we can not tell nothing more than these, which are still hidden mysteries.

It is really amazing and intuitive that detection and recognition are not independent and work together. We should try to recognize an object and the class information should be applied to segment the object and the segmentation result should be used to classify the object again and again. Does it sound familiar to you? Yes!!! It sounds like the EM algorithm. How beautiful!!!

- H. Choi

On-Orbit Calibration

D. T. Griffith, P. Singla, J. L. Junkins, "Autonomous On-Orbit Calibration Approaches for Star Tracker Cameras," AAS/AIAA Spaceflight Mechanics Meeting, No. 02-102, AAS, San Antonio, TX, Jan 27-30, 2002.

In addition to a batch calibration, we need on-orbit calibration because of the environmental changes over the life of the camera film. This paper describes the star tracker camera and proposes a way to estimate the distortion map of the star locations on-orbit (almost same word as the 'online' in machine learning), assuming the offsets and the focal length are given.
Basically they use the least square method with a couple basis function sets such as polynomials, sinusoidal functions or radial basis functions. By the health monitoring process, it updates the map only when the error is greater than a threshold.

It also includes a review of the least square method and the recursive least square method, which, by the way, looks like Kalman filter. The piecewise approximation is not an interesting part because usually the map would be smooth enough so that a few basis function might be good enough. And, this piecewise approximation needs a lot of basis functions which is computationally so expensive.


Nice paper for novices like me in the on-orbit calibration problem.

- H. Choi

Finding offsets and the focal length of a camera

M. A. Samaan, "Toward Faster and More Accurate Star Sensors Using Recursive Centroiding and Star Identification," Ph.D. Thesis, Texas A&M University, August 2003.

Chapter 3 is titled 'Ground Calibration for the Bore-Sight Offsets and the Focal Length.'

He uses the least square minimization (LSM) method recursively to find optimum offsets and the focal length which minimizes the error between inner product of reference stars and the estimated inner product of measured stars.

- H. Choi