Instead of Segmentation task, I am experimenting with PASCAL VOC data data to use it to train a `U^2 Net for Salient Object detection and ModeNet as Matting task.
Idea is to generate a B&W mask image from the coloured mask given (I'll take care of multi class and multi object thing). I can easily do it with 2 lines of code as:
img = np.array(Image.open('image1.png'))
img[img != 0] = 255
But the real difference between Matting / SOD
and Segmentation is that the boundaries for the given data are very rough unlike the boundaries of SOD and Matting. Is there a way to smooth the boundaries automatically for all the 3K images?
Here are the images, mask and converted B&W mask.