Please post your thoughts on the bag-of-words approaches that Alyosha will be talking abou tomorrow.

Edited by author 01-31-2006 03:19 PM
In Renninger's paper, can someone explain to me why 20 ms of jumbled scene mask needs to be shown? Why can't just the blank screen be shown immediately?

This paper is trying to say that texture recognition is used to identify picture during the early stage. Their reasoning is based on the positive correlation between the errors made by human and their model.
I disagree to their argument because, first, they're assuming that their texture histogram matching is optimal and will result in similar performance when human (also assumed to perform optimal texture matching) use only texture information. Their correlation measure is also odd. I don't understand why they compared correlation of errors for each category instead of correlation of each trials. They only had 10 categories for the basic-level and 3 categories for the superordinate-level to compute the correlation. Looking at the plots, it is not obvious to me that they are positively correlated. There were even 2 cases where it showed negative correlation and the authors got away with it by saying there is more than texture processing going on.

In Csurka's paper, I just wanted to point out an interesting opposing argument that the authors made against affine SIFT. The authors argued (in section 2.1) that (1) the world is 3D so affine transform is not enough and (2) increasing invariance will lose discriminative information. At first I thought it was a good observation against affine invariant detector, but then I realized that the two points were contradictory. One saying we need more parameters and the other saying we should use less parameters. Then it made me believe that affine-invariance was a good compromise.

My final words is that although Csurka et al claims that their method is novel, I can not find parts in their method which are novel. They have used standard affine-SIFT, standard k-means clustering, standard Naive-bayes & SVM. Maybe combining all those together and applying it to categorization?

About your question:
> In Renninger's paper, can someone explain to me why 20 ms of jumbled scene mask needs to be shown? Why can't just the blank screen be shown immediately?

Because they want to make sure that the image is really only available to the subject for the time that it is being shown. The noise mask is used to reduce the visual afterimage that your retina might carry for a few ms after the actual image turned blank. There are some fun psychology experiments where you stare at some pattern for 30s and then look at a white screen and you will still see a pattern for a few seconds, although the screen is really blank. The noise mask should "overwrite" that old visual stimulus.

Edited by author 01-31-2006 09:06 PM
 I agree with the Renninger paper to the fact that texture recognition aids in scene identification at an early stage. But it maybe unfair to say that texture alone is enough to perform a good match. Human scene identification seems to do much more than a simple texture match. A variety of low-level and high level cues such as color, shapes and semantic information go into recognizing an object. For starters, spatial frequency orientations seem to matter a lot to us in rapid scene categorization. As demonstrated in Schyns and Oliva (1994), rapid scene categorization seems to rely on estimating the coarse structure of the image first. A perfect example would be the "hybrid" images in Schyns and Oliva. It would be interesting to see how the texture recognition algorithm in the Renninger paper classifies these images.

On similar lines of the paper, one can wonder how important is color as a holistic cue in rapid scene categorization. As noted in Schyns and Oliva (2000), color can play a dual role; proper coloring can aid in rapid scene categorization and abnormal coloring can impair it. Also, color does not seem to matter for recognition in scenes where color does not convey the meaning of the scene (as in scenes of man made objects) whereas in natural scenes (such as beaches) color does play a significant role in rapid categorization of the scene. It maybe interesting to note that in the Renninger experiments, as the subjects' recognition task of outdoor scenes (beaches and forests) was impaired due to lack of color as a cue, the texture classifier probably performs as good as the subject in the natural/outdoor category.

Scene identification without localizing any objects -- whether it be using Torralba's gist descriptor or some texton histogram signature -- is important if one wants to use
top-down information when recognizing objects in images.

Although when used by themselves, such holistic techniques produce only decent scene classification results and appear not to be too useful for localizing objects in images.

However, if one
wants to avoid scanning a large number of templates (more objects implies more templates) across an image, then one must use a holistic descriptor like Renninger's texton histogram or Torralba's spectral signature in order to determine what objects to look for first and where to look for them.

Just a small comment about the Criminisi,Minka paper. The main contribution of this paper is automatically deciding on the dictionary size. They start out with a big dictionary, and keep mergins the keypoints (textons) in a greedy fashion. They claim that the resulting dictionary size is optimal.

Just wondering if greedy approach would indeed be optimal in this case? If not, is there a way to come up with an approximation bound for the greedy approach, or can there be cases when it can be arbitrarily bad?

Didn't find any inflatables what you're looking for? Not only will you benefit from the same quality and assurance found in our manufacturing of commercial inflatable units, but you will also be working with the "best of the best" in design and concept professionals in for the inflatables industry. Now welcome to do business with us, we provide Christmas Inflatables, Inflatable Jumper, Inflatable Toys and so on.

zelerborod