low impact of GC pairwise regul.
Borda opened this issue · comments
Description
Investigate an issue with nearly zero impact of the pairwise term in GraphCut segmentation on small elements with high confidence belonging to one class. Even these peaces are composed only a few superpixels, and the regularisation is set on maximum, they do not disappear...
import os
from skimage import segmentation, io
import matplotlib.pylab as plt
import imsegm.utils.data_io as tl_io
import imsegm.pipelines as seg_pipe
name = 'Lh05-04'
PATH_IMAGES = os.path.join(tl_io.update_path('data_images'), 'langerhans_islets')
img = io.imread(os.path.join(PATH_IMAGES, 'image', name + '.jpg'))
annot = io.imread(os.path.join(PATH_IMAGES, 'annot', name + '.png'))
# plt.imshow(img), plt.contour(annot, colors='b')
img_red = img[:, :, 0]
SLIC_SIZE = 25
SLIC_REGUL = 0.2
DICT_FEATURES = {'color': ['mean', 'median']}
model_seg, list_slic, _, _ = seg_pipe.train_classif_color2d_slic_features([img_red], [annot], DICT_FEATURES, sp_size=SLIC_SIZE, sp_regul=SLIC_REGUL)
plt.figure()
plt.imshow(segmentation.mark_boundaries(img[800:900, 1100:1300], list_slic[0][800:900, 1100:1300], color=(1, 1, 1)))
segm_gc, seg_soft = seg_pipe.segment_color2d_slic_features_model_graphcut(img_red, model_seg, DICT_FEATURES, sp_size=SLIC_SIZE, sp_regul=SLIC_REGUL, gc_edge_type='ones', gc_regul=5000000.)
plt.figure()
plt.imshow(segm_gc), plt.contour(annot, colors='g'), plt.title('segmenatation')
plt.figure(figsize=(10, 3))
plt.subplot(1, 2, 1), plt.imshow(seg_soft[800:900, 1100:1300, 0]), plt.title('class 0'), plt.colorbar()
plt.subplot(1, 2, 2), plt.imshow(seg_soft[800:900, 1100:1300, 1]), plt.title('class 1'), plt.colorbar()
Way to reproduce
- Code example
- Relevant images
- packages version (gco-wrapper==3.0.3)
Probably investigate some numerical instability in GraphCut, compute the global energy with and without assimilated the small region (bottom right).
"Have you looked at opengm? I'm pretty sure they have their own implementation. Opencv does, too..."
amueller/gco_python#18
Have look at https://networkx.github.io
probably can be solved by #21