Hello, When looking into your code, I find that you used the SBDD dataset for training(according to your prototxt, right?). However, it's expressed in the original paper that the author just used the augmented data to train(see below). So could this be a reason for your failure to achieve the results declared in paper?

Since the labeling of augmented data is not very accurate, we use the full set of 11,685 images to train the OBP-FCN-32s only.

What's more, I've check your layers.py and observe that your load_contourlabel function seems to not follow the idea in the original paper.

   def load_contourlabel(self, idx):
        """
        Load label image as 1 x height x width integer array of label indices.
        The leading singleton dimension is required by the loss.
        """
        import scipy.io
        mat = scipy.io.loadmat('{}/cls/{}.mat'.format(self.sbdd_dir, idx))
        label = mat['GTcls'][0]['Segmentation'][0].astype(np.uint8)        
                   
        from scipy import ndimage
        edge_horizont = ndimage.sobel(label, 0)
        edge_vertical = ndimage.sobel(label, 1)
        magnitude = np.hypot(edge_horizont, edge_vertical)     

        contour_label = np.zeros(label.shape,dtype=np.uint8)
        contour_mask = np.zeros(label.shape,dtype=np.uint8)
        label[label==0]=254
        label[label<254]=1
        
        contour_mask[magnitude>0] = 255            
        kernel = np.ones((5,5),np.uint8)
        import cv2
        #contour_mask = cv2.dilate(contour_mask,kernel,iterations = 1)
        #contour_mask = cv2.erode(contour_mask,kernel,iterations = 1)
        contour_label[label==1]=1   
        contour_label[label==254]=0 
        contour_label[contour_mask==255]=2

        contour_label = contour_label[np.newaxis, ...]
        return contour_label

Because according to the Algorithm A for Label Conversion for the PASCAL VOC dataset and the corresponding sentences,

your order of assignment of label is different from the paper, while the author said

Note that the order of the assignment is very important so as to keep the completeness of object and the accuracy of the boundary

I think the order of the assignment matters too.
Last, I am not very sure your strategy to find the boundary has follows the author's, since theirs adopt a 3 × 3 window for comparison, which I fail to find the action in your code.

:)

Hello, you said that the original paper claimed a 87.5% mean IU in PASCAL VOC 2011 validation set. Are you serious? 87.5% is a unbelievably high score.

I actually find they stated "the OBG-FCN with w = 2 reaches 69.5% mean IU in VOC 2011 test and 69.1% in VOC 2012 test, outperforming the baseline FCN by about 7%."

So I think it's probably a typo

:)

Hello,

where did you got these .prototxt and .caffemodel files?

for example look at this layer at the FCN-8s-OBG-4s.

layer {
  name: "obg-fcn-fuse"
  type: "Eltwise"
  bottom: "mask_projection_no_b"
  bottom: "score_fcn"
  top: "score"
  eltwise_param {
    operation: PROD
  }
}

where this top: "score" refers to? there is not layer in the name of score in the whole text. I think the problem of lower accuracy might come from this or possible similar problems inside prototxt files.