compute histogram

[siggie]

Hey,
I dont't get it what to do exactly when computing the histogram. Do I have to compute and store a histogram for each training image or do I have to compute the histograms only for the two classes 'plane' and 'bike' which sums up all occurances of a codword over all images of a class. In the latter case do we need to normalize the histogram?

Thanks for your help!

[ampelmann]

I just computed histogram-values for each image. with 120 images, i have an output-histogram 120x50. for example in histogram(20,10), you find the value how many feature vectors of image 20 were assigned into cluster 10. That's how I understood it and I really hope it's not totally wrong cause I am already getting insane with this homework.

Hope that helps

PS: Read the first sentence of Part 2.B. I guess that's where these histogram is needed for...

[lustiz]

As is turns out there are two steps which involve computing a histogram. However, we are actually only concerned with one of them.

1) First, sift features are nothing else than histograms over the gradient discretized by a certain number of directions (this is already done by the given implementation).

2) Second, the features we are using for classification are NOT the sift features. Prof. Roth really put a lot of focus on this point. Again, we are NOT using sift features directly. Instead, we compute a codebook over all sift feature vectors (KMeans). Having finished that, we then compute the actual features as a histogram by finding the closest visual codeword in the codebook for each feature vector (corresponding to each interest point) in a given image. In other words: We count how often certain visual words occur in an image. The resulting histogram PER IMAGE serves as input to the classifiers (normalized in the case of the svm).