Very poor results with MulticlassSupportVectorMachine (BagOfWords), MulticlassSupportVectorLearning
rmk60
I try to classify grayscaled images of size range 40x40..64x64. The images are birds, airplanes, copter and clouds for first. (Train data = 360, test data = 200)
But the classify results are very poor < 50%!
Approx. half of all images have a feature vector of 0. So when I remove all these images, result will be slightly better but unfortunately min. 1 category will be lost (empty).
I use SURF-Detector and tried Chi-Square, HOG and Gauss Kernel. Also varying the numer of words in range 10..100 will have no positive effect.
- Why is the feature extraction so poor though pictures are of good quality?
- Why has the complexity of the learning algorithm no influence on training results?
My goal was to resize the images to 32x32 but in this case images have empty feture vectors. Is there any limit concerning the image size?
Thank you in advance for any help.