If someone could try to briefly explain pLSA and LDA in simpler terms I'd appreciate it.

Alright, so I get a really easy comment this time. I HIGHLY recommend reading Ian Fasel's thesis at http://mplab.ucsd.edu/~ianfasel/IanThesis.pdf Similar content, VERY different approach and viewpoint.

The paper talks a lot about large sets of images. Most large sets of images currently is found on the internet (and that set is growing). However, the vast majority of image data today is accompanied by human labeling. I think that as a follow up to this paper, text labels should be incorporated into the process of learning classes of objects. Clearly if images on the web were perfectly labeled, there would be no need for this kind of research. However, the fact that text exists with images in almost every case (thanks to the web), cannot be ignored.

The paper chose Normalized Cuts for its segmentation algorithm, because the size of the segments were similar to the size of possible objects. Could you give a quick overview of Normalized Cuts and maybe some top competitors that could have been used instead...

A main point of the paper seems to be the use of the multiple segmentations. Yet the results don't seem to show that the multiple segmentations provide that much of a benefit. Why would the single-segmentation approach outperform in the top 20 returned images (it doesn't over the top 500)?

This paper is a clear rip off of the seminal work by Ben-Haim et al. (http://www.cs.ucsd.edu/~sjb/slam06.pdf).

In all seriousness though, I think it would be really cool if this type of approach was applied to image search engines.

Edited by author 10-12-2006 01:31 AM
matt: I was looking at their website and they seem to have given images for all the topics at http://people.csail.mit.edu/brussell/resea...iscovery/index.html

I am bit confused by the results they show on cars (Figure 6, top-left), in which a black and white car falls under same topic. This is not possible by using visual words (in color or b/w images) if the descriptor has any sort of color or intensity information.

btw do they use color images in their dataset or b/w?

In figures 5-7, I'm somewhat troubled by the fact that they've selected a handful of the topics out of the set of algorithm-discovered topics. It makes me wonder about the topics they didn't select.

I also wonder whether the algorithm could be adapted to partly labeled data (where (some of) the contents of a picture are labeled, but no knowledge of where they occur in the image is provided). Since it doesn't seem like this method is likely to scale well to cases where the data set isn't already set up to have a small selection of objects, it seems like these labels would already be available. (There isn't as much of a correlate in the text-mining community for this kind of approach, but...)