Hi,
I'm developing a flower detector and would be glad if anyone has some ideas I could try.
Current directions I was thinking of:
Any other directions you can suggest?
Thanks in advance, Gil.
Hi, as StevenPuttemans already answered, contains your list the most frequent used techniques. But you should also take a look at the Bag-of-Words approach. It is widely used for image classification and object categorization. There exist a project at Visual Geometry Group from Andrew Zisserman with the goal to recognize the species of flowers using bag-of-words. On the project website you can find publications, a dataset and test your own images by uploading the image file.
Hi, LBP+SVM is also a possible technique. On the PASCAL VOC'06 object detection challenge the results of LBP+SVM and HOG+SVM are similar (see for example: http://hal.archives-ouvertes.fr/docs/00/69/56/27/PDF/lqp-tr-hal.pdf).
I am not an expert in BOW but I guess that the bag-of-words approach is less suitable to detect object (flowers) in an image. BOW computes and counts features in the whole image and compares them with a model. No spatial relation of the features is stored. So you have no information where the features which best match the flower-model are located in the image. As a conclusion i think that if you already have a hypothesis (region-of-interest) and you want to know if it is a flower you can use BOW. But for detecting the ROI of the flower in the image BOW is not suitable. In the cited project (http://www.robots.ox.ac.uk/~vgg/research/flowers_demo/) they use Grabcut to segment the flowers from the background and use then BOW for classification.
Just a side-note: localization is typically important for an instance classification via BoW. It can be easily achieved by using a spatial pyramid, i.e. a form of grid and compute in each grid a BoW-descriptor which are in the end concatenated to one descriptor. See also http://www.cs.illinois.edu/homes/slazebni/publications/cvpr06b.pdf .
Thank you all for your answers.
My first attempt is taking each patch in the image, re sizing it to 400X400 pixels, then dividing it into 4 100X100 patches and extracting rotation invariant uniform LBP codes. These are fed to a previously trained SVM.
Current results are far from satisfactory, but I'm working on tuning the SVM to get better performance.
Any comment on what I've done so far?
Thanks!
Asked: 2013-08-05 12:20:28 -0600
Seen: 3,064 times
Last updated: Aug 06 '13
I think you hitted the most frequent used techniques for object detection: You could look into class specific randomn forests which is sortlike svm.