Can anyone suggest for me the best image features for scene classification. I have programmed with SURF, SIFT, BoVW, Kernel Map,the GIST feature, Dense SIFT, Dense SIFT in a Gaussian Pyramid; Dense SIFT with spatial pooling (commonly known as PHOW - Pyramid Histogram of Words, and some tiny image features.
I have tried lots of awesome classifiers successfully.
Now, I want to move some steps further. Currently I want to experiment with non-linear support vector machine (SVM) and I am in need of suggestions for the best feature for ROBOST and INVARIANCE in scene recognition and object detection.
Every suggestion is appreciated.
Thanks
Hi there!
It is quite imprecise to claim that a specific feature will be robust and invariant across all sorts of scenes. Nevertheless, one predominant feature which gained much weight in the computer vision domain is the Histogram of Oriented Gradients (HOG) descriptor and which I see is missing from the list of features you experimented on.
Try that while waiting for other RGaters to input their suggestions.
Good luck!
@Abbas, thank you so much.
I am going to research on it and try it out.
I guess all current approaches for scene classification are based on convolutional neural networks. Is there any reason you are not considering these?
For some more ideas, I'd suggest going through the papers which cite a scene classification dataset like https://scholar.google.de/scholar?cites=18419800374042180760&as_sdt=2005&sciodt=0,5&hl=de
I would suggest using HOG with SVM, also as suggested, try CNNs since they have gained much traction in the ImageNet challenge recently
@Marcel and Vikas, Thank you so much, I got 97% prediction accuracy with CNN.
- Badges
- Science topic