Hello ı want to ask you one question about bag of words using SIFT feature. İf my train samples has the same number of instances the result is very high. Especially sıft and surf. because sıft and surf histogram used bag of words metod. and if I increase the number of cluster or increase my image size, My result will increase.Why the result changes with the number of instances or image size or number of cluster?Please help meee
Simply because you get a different (better or worse) background model to describe your data...
What do you mean by "very high result", or "my result will increase"? What do you consider as a "result" that "increases"?
accurate classification results or true classification results increase
By increasing the number of clusters, feature resolution (hence correct classification probability) would certainly increase. I am still not sure about the image size. Normally, you cannot increase the image size, anyway. But, if you use lower resolution versions and compare them to the "original" high resolution, then, of course the normal high resolution image will have better detail for the salient points. In fact, SIFT would then use both the lower resolution data, and the original resolution (due to being multi-resolution) - and that would increase the information content for better classification.
thank you very much your reply. but ıI don 't understand very well what do you mean "future resolution". I think correct classification probability increase because sift features normally distributed to cluster. A class of sift feature isn't dominant to another class sift feature.
The classification performance of BOVW model can be influenced by several factors, such as the sampling way(e.g. the size of sampling window, the sliding step size), the size of dictionary/the number of clusters, the number of training samples. Generally, the training samples are used for sift feature extraction, visual dictionary generation, and the classifier learning and construction. For all the factors, different parameters may result in different result. Besides, the k-means clustering itself is random, different initial centers will lead to different results.
I know boww model influenced several factors. I want to learn why influence?
Changing number of clusters (codebook size) or instances mostly affect the classifiers rather than image descriptors. For example, when you use more clusters (bigger codebook) you will end up with bigger feature vectors. Probably, there is more information in a feature vector with size 200 rather than 50. So, your classifier can classify these feature vectors more accurately. Same thing happen when you change number of instances; The more instances you have in the training phase, the more accuracy you can get.
Changing the image size is the only parameter that affects the image descriptor. When you increase the image size, sift can find more interest points. As a result, it will produce a richer descriptor for the images.
hi if your number of cluster are more means samples not close and the intra-class variations are high. So capturing the more difference or working on high range compared to small number of clusters. So you training samples capture more information compare to small number of clusters. This may be the strong reason for your results.
If you want to get more info please read the attached paper
http://link.springer.com/chapter/10.1007%2F978-81-322-1602-5_60
current histogram binning methods like Scott's normal reference rule or Freedman - Diaconis rule contains sample size as parameter. A new approach is developed, which calculates histogram bin depending on percentiles
Article PERCENTILE BASED HISTOGRAM BIN WIDTH
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Computer Software