seg-crf 1.0.0

Conditional Random Field Implementation for segmentation models as used in Deeplab-v2

Fully Connected CRF

This repo implements CRF as described in Deeplab paper it takes about 0.2 seconds per image. Following image is taken form DeepLab paper

Requirements

Python <= 3.6
pydensecrf
cv2
matplotlib
gray2color

It takes following arguments. For details visit project page.

        ⚠ Zero pixels are consdered background
        img_path : path to an image, 
                        Format [H, W, 3]; values ranging from [0, 255]
        model_op_path : path model output of the same input image.
                        Format [H, W]; values ranging from [0, num_of_classes]
        num_of_classes : number of classes in a dataset e.g. in cityscape has 30 classes
        clr_op : color the output or not a bool
        pallet2use : see https://pypi.org/project/gray2color/ for details
        img_w : for resizing image and mask to same size default is 1024
        img_h : for resizing image and mask to same size default is 512
        apperance_kernel : The PairwiseBilateral term in CRF a list of values in order [sxy, srgb, compat]  
                            default values are [8, 164, 100]
        spatial_kernel : The PairwiseGaussian term in CRF a list of values in order [sxy, compat]  
                            default values are [3, 10]

from seg_crf import Seg_CRF

img_path='D:/Anaconda/Image_analysis/cat.png'
model_op_path='D:/Anaconda/Image_analysis/mask.png'

crf = Seg_CRF(img_path, model_op_path, 2, img_w=1024, img_h=512, clr_op=True, pallet2use ='cityscape')

gray, rgb = crf.start()
plt.imshow(rgb)

Appearance and Spatial Kernel

# Default Values are
apperance_kernel = [8, 164, 100] # PairwiseBilateral [sxy, srgb, compat]  
spatial_kernel = [3, 10]         # PairwiseGaussian  [sxy, compat] 

# or if you want to to specify seprately for each XY direction and RGB color channel then

apperance_kernel = [(1.5, 1.5), (64, 64, 64), 100] # PairwiseBilateral [sxy, srgb, compat]  
spatial_kernel = [(0.5, 0.5), 10]                  # PairwiseGaussian  [sxy, compat] 
# Use like
crf = Seg_CRF(img_path, model_op_path, 2, img_w=1024, img_h=512,
                 apperance_kernel=apperance_kernel, spatial_kernel=spatial_kernel,
                 clr_op=True, pallet2use ='cityscape')

gray, rgb = crf.start()