zqh-sail / spx_svm

SVM classification on Superpixel

SPX_SVM

SVM classification on Superpixel.

Input : TrainImage, ROI and TestImage

Output : Superpixel of TestImage classified by SVM, separating foreground superpixels from background superpixels

Detail :

1. Training

• Draw ROI on TrainImage to select Foreground region
• Apply grabCut (from openCV) to have coarse fgnd/bgnd subtraction
• Oversegment TrainImage with SLIC to get the superpixels
• Select foreground superpixel samples if #pixel_fgnd>threshold inside that superpixel
• Select background superpixel samples in area around ROI
• Do SVM training with those samples

2. Test :

• Segment TestImage in Superpixel
• Extract feature from each Superpixel
• Classify each Superpixel with SVM

Note : Different settings are available

• Color Space : BGR, HSV, Lab
• Feature : MEAN_COLOR, HISTOGRAM3D
• SVM Types : ONE CLASS , TWO CLASS
• SVM Kernel : LINEAR, RBF
• SLIC parameter : size/number of Superpixel, compactness
• Background ROI size : fullFrame -> all the superpixel out of ROI, or scaleROI -> BROI = X*ROI-ROI