Regarding to "Object Recognition with Informative Features and linear classification".
This method of trying to detect objects using fragments of it reminds me eigen-based methods. It is a similar approach in which instead of using fragments of the object usually a small set of bases are used to represent all possible views of the entire object. These methods usually perform better than simple cross-correlation as tend to eliminate outliers and are faster since fewer basis are needed to encode the object. I would like have seen a comparison of this method with eigen methods.

Edited by author 02-01-2006 09:57 AM
I'm not concerned much with the percentages of their results, but I think that it's a bit suspicious when they imply they've achieved some computational benefit for the classification stage by using informative features rather than simple ones. In essence, they're simply shifting complexity from classification to feature extraction. The decision surface may be linear in the classification stage, but their informative feature templates still model local conditional dependence relationships between pixels. It's not surprising that the TAN classificiation would be redundant in that case. But in the end it's not clear that an "informative feature" strategy would scale any better than a simple feature strategy would.

Although I agree that the dataset this paper uses is a bit limited, the main challenge that is underlying is the resolution of the images. I think that due to the low resolution of the objects within the image and given the specific application that they are addressinng, the common approaches (wavelets, etc.) don't necessarily perform their best. I'm not too sure how well other approaches that use similar size subwindows might perform in this dataset but, I'd guess that they would suffer some decrease in performance as well. Since not all the features observed in a subwindow of a high res image is captured within a low res image of equal size. Similarly, I also think that applying this approach on higher resolution images might produce worse results (or a result similar to a pure brute force search algorithm) than the common approaches.

I agree that good descriptors are more important than good classifiers. I also had the same impression on my own work.

But I'm confused a bit... I may be misunderstanding something, but aren't the standards these days something like > 90% detection and < 10^-4 false positives for each subwindow with window based detection? I think the task here to classify 14x21 is similar to classifying a subwindow.

I guess this paper deliberates on the trade off between a better feature representation and a better classifier for efficient recognition. Results indicate that the fragment based approach is better than the wavelet approach irrespective of the kind of classifier that is chosen. The authors stress that the performance of wavelet based classifiers could be increased but only at the cost of heavy computations. The authors seem to indicate that a better choice of features has precedence over a good choice of classifiers. However, as Pete indicated the dataset may be insufficient and also wavelets are not the only techniques for generic feature extraction?

Edited by author 01-31-2006 03:39 PM
Vidal and Ulman use just low-res pictures of cars for their '03 paper, and I wonder if they're getting such good results just because they've created a spanning set of the side view of cars. They started with a set of 59 thousand fragments, and I wonder if the results were that much better than an exhaustive search would be. On the other hand, maybe using a heuristic on a huge amount of data is what people do too, and it might be a very robust classifier. But still, I'd like to see them try to correctly classify several different types of objects, rather than deciding if a query is inside or outside one class.

Please Posts your thoughts on the papers by Ullman & Vidal-Naquet which I will be presenting tomorrow.