Please post your thoughts on the paper "Pedestrian Detection in Crowded Scenes" here.

The ROC curves look pretty good, but I can't figure out what happened to the red one. It seems to just quit at .3 precision. Also, I'm a little suspicious of the photo results, as they seem to be able to get a pretty good contour on the people, which we've seen to be a really difficult task even when that is the primary goal. Lastly, I'm a little mixed up about how they chose their kernel in the training section, maybe it has something to do with the balloon density estimators?

Edited by author 04-05-2006 02:44 AM
I am not so sure about this, but the reason they get good contours might be because of the small solution space they search.

I find it interesting that the contour they fit to the example image is one of the 210*2 contours in the training image, yet the fit looks good. It is suggesting that all possible silhouette of pedestrians are one of 210*2. That is a pretty small search space. Compare it with this: say the contour is modeled by 5 eigenvectors obtained from PCA and the weight for the vectors are discretized into 4 values. That's already 4^5=1024 possible solutions. Usually the weights aren't discretized and there are more than 5 eigenvectors. (Recall face AAM)

I'm not sure if such method is good or bad. In fact, this may be the weakest part in this paper. I just thought it was worth mentioning.

basbasd