An numpy-based implement of PointRend

This is an implement a PointRend function for Segmentation result refinement. The paper can be find at https://arxiv.org/pdf/1912.08193.pdf

copy the pointGenerate.py to your directory and you are ready to rock.

from pointGenerate import getpoint
my_mask = np.asarray(Image.open("tree_mask.jpg").resize((32,32)))
# convert this 3-channel binary mask to a 1-channel binary one
my_mask = my_mask[:,:,0]
# get the point, nearest_neighbor chose the sample points locations
points = getpoint(my_mask, k=2, beta = 0.95, nearest_neighbor=1)

# plot the result
points = list(zip(*points))
plt.imshow(my_mask,cmap="Purples")
plt.scatter(points[1],points[0],c='black',s=4)

the original image and mask:

when the mask is 32*32

when the mask is 64*64

when the mask is 128*128

When I was using this, I find the speed is horrible , so I improved the point selection process by storing the it. The point selection process is significantly accelerated while the image is large. However you want to use the old one, just use getpoint(new_if_near=False). The performance improvement is shown below, and a fancy but totally unnecessary figure is plotted.

1. We will test this on V-Net to see if there are some segmentation performance boost.
2. If any one feel like trying the PointRend on Mask-rcnn, leave the comments in the issue.