def pixel_unary(imageset, filepattern):
if isinstance(imageset, ds.DatasetImage):
fn = filepattern.format(name=imageset.name, dspath=imageset.ds.path)
pixel_unaries = load_probimage(fn)
imgdim = imageset.top_masks.shape[:2]
assert imageset.ds.classnum == pixel_unaries.shape[-1]
if pixel_unaries.shape[:2] != imgdim:
if pixel_unaries.shape[1::-1] == imgdim:
pixel_unaries = np.transpose(pixel_unaries, axes=(1, 0, 2))
else:
logging.warn("Dimension mistach for dense unaries of {}".format(imageset) )
return pixel_unaries
else:
return np.hstack((pixel_unary(n, filepattern=filepattern) for n in imageset))
def region_unary_from_probfile(img, filepattern):
fn1 = filepattern.format(name=img.name, dspath=img.ds.path)
pu = load_probimage(fn1)
reg = img.regions
regids = np.unique(reg)
regids.sort()
assert(np.all(regids == np.arange(len(regids))))
if pu.shape[:2] != reg.shape and pu.shape[:2] == reg.shape[::-1]:
pu = np.swapaxes(pu, 0, 1)
res = np.zeros((len(regids),img.ds.classnum))
for i in xrange(pu.shape[0]):
for j in xrange(pu.shape[1]):
res[reg[i,j],:] += pu[i,j,:]
res /= np.sum(res, axis=1)[:, np.newaxis]
return res
def unary_from_probfile(imageset, filepattern, with_ids=False):
""" contructs pS from densecrf probimages"""
if isinstance(imageset, ds.DatasetImage):
fn1 = filepattern.format(name=imageset.name, dspath=imageset.ds.path)
pixel_unaries = load_probimage(fn1)
masks = imageset.top_masks
unary = np.empty((imageset.ds.classnum, masks.shape[-1]))
for i in range(masks.shape[-1]):
unary[:,i] = pixel_to_segment_prob(masks[:,:,i], pixel_unaries)
assert np.allclose(np.sum(unary,axis=0), np.ones_like(np.sum(unary,axis=0)), rtol=1e-2)
unary /= unary.sum(0)
if with_ids:
return unary, range(unary.shape[1])
else:
return unary
else:
return np.hstack((unary_from_probfile(n, filepattern=filepattern,with_ids=with_ids) for n in imageset))
