def output_debug(self,kp0,ownership,iterstring,weights=None):
try:
from mylib import viz
except:
return
n_models = self.models.shape[0] #self.params['n_models']
n_coeffs = self.flow_bases_u.shape[0]
sz_ownership = ownership.shape[0]
if sz_ownership < n_models:
z = np.zeros((n_models-sz_ownership,ownership.shape[1]))==1
ownership = np.vstack((ownership,z)).copy()
flow_u_all = np.zeros((n_models,self.flow_bases_u.shape[1]))
flow_v_all = np.zeros((n_models,self.flow_bases_u.shape[1]))
for m in range(n_models):
flow_u_all[m] = self.models[m,:n_coeffs].dot(self.flow_bases_u)
flow_v_all[m] = self.models[m,n_coeffs:].dot(self.flow_bases_v)
if weights is None:
u_combined,v_combined = self.get_flow_nearest(kp0,ownership)
else:
u_combined,v_combined = self.get_flow_nearest(kp0,weights)
I_combined = viz.viz_flow(u_combined,v_combined)
I_individuals = []
for m in range(n_models):
I_individuals.append(viz.viz_flow(
flow_u_all[m].reshape((self.pc_h,self.pc_w)),
flow_v_all[m].reshape((self.pc_h,self.pc_w))))
n_x = int(np.ceil((n_models+2)/3.0))
n_y = 3
colors = np.argmax(ownership,axis=0)
if have_plt:
plt.figure(figsize=(n_x*8,n_y*4))
plt.subplot(n_y,n_x,1)
plt.scatter(kp0[:,0],kp0[:,1],linewidths=0,s=10,c=colors,vmin=-1,vmax=n_models,cmap='cubehelix')
plt.xlim([0,self.pc_w])
plt.ylim([self.pc_h,0])
plt.title('Point ownerships')
plt.subplot(n_y,n_x,2)
plt.imshow(I_combined)
plt.title('Combined flow')
for m in range(n_models):
plt.subplot(n_y,n_x,3+m)
plt.imshow(I_individuals[m])
inds = ownership[m]
plt.scatter(kp0[inds,0],kp0[inds,1],linewidths=0,s=10,c='black')
plt.xlim([0,self.pc_w])
plt.ylim([self.pc_h,0])
plt.title('Flow model {0}'.format(m))
plt.savefig('./output_{0}.png'.format(iterstring),dpi=100,bbox_inches='tight')
plt.close()
def output_debug(self, kp0, ownership, iterstring, weights=None):
try:
from mylib import viz
except:
return
n_models = self.models.shape[0] #self.params['n_models']
n_coeffs = self.flow_bases_u.shape[0]
sz_ownership = ownership.shape[0]
if sz_ownership < n_models:
z = np.zeros((n_models - sz_ownership, ownership.shape[1])) == 1
ownership = np.vstack((ownership, z)).copy()
flow_u_all = np.zeros((n_models, self.flow_bases_u.shape[1]))
flow_v_all = np.zeros((n_models, self.flow_bases_u.shape[1]))
for m in range(n_models):
flow_u_all[m] = self.models[m, :n_coeffs].dot(self.flow_bases_u)
flow_v_all[m] = self.models[m, n_coeffs:].dot(self.flow_bases_v)
if weights is None:
u_combined, v_combined = self.get_flow_nearest(kp0, ownership)
else:
u_combined, v_combined = self.get_flow_nearest(kp0, weights)
I_combined = viz.viz_flow(u_combined, v_combined)
I_individuals = []
for m in range(n_models):
I_individuals.append(
viz.viz_flow(flow_u_all[m].reshape((self.pc_h, self.pc_w)),
flow_v_all[m].reshape((self.pc_h, self.pc_w))))
n_x = int(np.ceil((n_models + 2) / 3.0))
n_y = 3
colors = np.argmax(ownership, axis=0)
if have_plt:
plt.figure(figsize=(n_x * 8, n_y * 4))
plt.subplot(n_y, n_x, 1)
plt.scatter(kp0[:, 0],
kp0[:, 1],
linewidths=0,
s=10,
c=colors,
vmin=-1,
vmax=n_models,
cmap='cubehelix')
plt.xlim([0, self.pc_w])
plt.ylim([self.pc_h, 0])
plt.title('Point ownerships')
plt.subplot(n_y, n_x, 2)
plt.imshow(I_combined)
plt.title('Combined flow')
for m in range(n_models):
plt.subplot(n_y, n_x, 3 + m)
plt.imshow(I_individuals[m])
inds = ownership[m]
plt.scatter(kp0[inds, 0],
kp0[inds, 1],
linewidths=0,
s=10,
c='black')
plt.xlim([0, self.pc_w])
plt.ylim([self.pc_h, 0])
plt.title('Flow model {0}'.format(m))
plt.savefig('./output_{0}.png'.format(iterstring),
dpi=100,
bbox_inches='tight')
plt.close()
def debug_GC(self,
unaries_warp,
unaries_colors,
unaries_dist,
tot,
res,
w_y,
w_x,
flow_u_all=None,
flow_v_all=None):
try:
from mylib import viz
except:
return
n_models = unaries_warp.shape[2]
n_coeffs = self.flow_bases_u.shape[0]
n_pixels = self.flow_bases_u.shape[1]
I_warp = np.vstack([unaries_warp[:, :, m] for m in range(n_models)])
I_col = np.vstack([unaries_colors[:, :, m] for m in range(n_models)])
I_dist = np.vstack([unaries_dist[:, :, m] for m in range(n_models)])
I_tot = np.vstack([tot[:, :, m] for m in range(n_models)])
#I_tot = I_warp + I_col + I_dist
if have_plt:
plt.figure(figsize=(10, 10))
plt.subplot(1, 4, 1)
plt.imshow(I_warp, cmap='hot')
plt.title('Warp unaries')
plt.subplot(1, 4, 2)
plt.imshow(I_col, cmap='hot')
plt.title('Color model unaries')
plt.subplot(1, 4, 3)
plt.imshow(I_dist, cmap='hot')
plt.title('Distance unaries')
plt.subplot(1, 4, 4)
plt.imshow(I_tot, cmap='hot')
plt.title('Combined unaries')
plt.savefig('unaries.png', dpi=200, bbox_inches='tight')
plt.close()
if res is not None:
plt.figure(figsize=(10, 10))
plt.imshow(res, cmap='hot')
plt.title('Model results')
plt.savefig('result_models.png', dpi=200, bbox_inches='tight')
plt.close()
plt.figure(figsize=(10, 10))
plt.subplot(2, 1, 1)
plt.imshow(w_y, cmap='hot')
plt.colorbar()
plt.title('Vertical regularization weights')
plt.subplot(2, 1, 2)
plt.imshow(w_x, cmap='hot')
plt.colorbar()
plt.title('Horizontal reg weights')
plt.savefig('reg_weights.png', dpi=200, bbox_inches='tight')
plt.close()
plt.figure(figsize=(10, 10))
plt.subplot(411)
plt.imshow(np.argmin(unaries_warp, axis=2), cmap='hot')
plt.title('min of warp unaries')
plt.subplot(412)
plt.imshow(np.argmin(unaries_colors, axis=2), cmap='hot')
plt.title('min of color unaries')
plt.subplot(413)
plt.imshow(np.argmin(unaries_dist, axis=2), cmap='hot')
plt.title('min of dist unaries')
plt.subplot(414)
plt.imshow(np.argmin(unaries_warp + unaries_colors + unaries_dist,
axis=2),
cmap='hot')
plt.title('min of combined unaries')
plt.savefig('unaries_min.png', dpi=200, bbox_inches='tight')
plt.close()
if flow_u_all is not None:
plt.figure(figsize=(10, 10))
nx = 3
ny = int(np.ceil(float(n_models) / 3.0))
for n in range(n_models):
u = flow_u_all[n].reshape((256, 512))
v = flow_v_all[n].reshape((256, 512))
Iv = viz.viz_flow(u, v)
plt.subplot(ny, nx, n + 1)
plt.imshow(Iv)
plt.title('Model {}'.format(n))
plt.savefig('./models_all.png', dpi=200, bbox_inches='tight')
plt.close()
def output_debug2(self, kp0, point_models, res, flow_u_all, flow_v_all):
try:
from mylib import misc
except:
return
n_models = flow_u_all.shape[0] #self.params['n_models']
n_coeffs = self.flow_bases_u.shape[0]
# sz_ownership = ownership.shape[0]
# if sz_ownership < n_models:
# z = np.zeros((n_models-sz_ownership,ownership.shape[1]))==1
# ownership = np.vstack((ownership,z)).copy()
#
I_individuals = []
for m in range(n_models):
I_individuals.append(
viz.viz_flow(flow_u_all[m].reshape((self.pc_h, self.pc_w)),
flow_v_all[m].reshape((self.pc_h, self.pc_w))))
#colors = np.argmax(ownership,axis=0)
colors = flow_u_all.shape[0]
path = self.debug_path
if have_plt:
# Point ownershipts
plt.figure(figsize=(10, 10))
plt.scatter(kp0[:, 0],
kp0[:, 1],
linewidths=0,
s=40,
c=point_models,
vmin=0,
vmax=n_models,
cmap='cubehelix',
alpha=0.5)
#plt.axis('equal')
plt.xlim([0, self.pc_w])
plt.ylim([self.pc_h, 0])
plt.xticks([], [])
plt.yticks([], [])
ax = plt.gca()
ax.set_aspect('equal')
#plt.title('Point ownerships')
plt.savefig(os.path.join(path, 'ownerships.png'),
dpi=200,
bbox_inches='tight',
pad_inches=0)
plt.close()
for m in range(n_models):
plt.figure(figsize=(10, 10), frameon=False)
plt.imshow(I_individuals[m])
#inds = ownership[m]
inds = point_models == m
plt.scatter(kp0[inds, 0],
kp0[inds, 1],
c='black',
alpha=0.5,
linewidths=(1, ),
s=40)
#print(kp0[inds,0])
#print(kp0[inds,1])
plt.xlim([0, self.pc_w])
plt.ylim([self.pc_h, 0])
plt.xticks([], [])
plt.yticks([], [])
#plt.title('Flow model {0}'.format(m))
if m == n_models - 1:
outfname = 'model_homography.png'
elif m == n_models - 2:
outfname = 'model_pca.png'
else:
outfname = 'model_{0:02}.png'.format(m)
plt.savefig(os.path.join(path, outfname),
dpi=200,
bbox_inches='tight',
pad_inches=0)
plt.close()
plt.figure(figsize=(10, 10), frameon=False)
plt.imshow(res, cmap='cubehelix', vmin=0, vmax=n_models - 1)
#plt.title('Segmentation')
#plt.colorbar()
plt.xticks([], [])
plt.yticks([], [])
plt.savefig(os.path.join(path, 'segments.png'),
dpi=200,
bbox_inches='tight',
pad_inches=0)
plt.close()
def output_debug2(self,kp0,point_models,res,flow_u_all,flow_v_all):
try:
from mylib import misc
except:
return
n_models = flow_u_all.shape[0] #self.params['n_models']
n_coeffs = self.flow_bases_u.shape[0]
# sz_ownership = ownership.shape[0]
# if sz_ownership < n_models:
# z = np.zeros((n_models-sz_ownership,ownership.shape[1]))==1
# ownership = np.vstack((ownership,z)).copy()
#
I_individuals = []
for m in range(n_models):
I_individuals.append(viz.viz_flow(
flow_u_all[m].reshape((self.pc_h,self.pc_w)),
flow_v_all[m].reshape((self.pc_h,self.pc_w))))
#colors = np.argmax(ownership,axis=0)
colors = flow_u_all.shape[0]
path = self.debug_path
if have_plt:
# Point ownershipts
plt.figure(figsize=(10,10))
plt.scatter(kp0[:,0],kp0[:,1],linewidths=0,s=40,c=point_models,vmin=0,vmax=n_models,cmap='cubehelix',alpha=0.5)
#plt.axis('equal')
plt.xlim([0,self.pc_w])
plt.ylim([self.pc_h,0])
plt.xticks([],[])
plt.yticks([],[])
ax = plt.gca()
ax.set_aspect('equal')
#plt.title('Point ownerships')
plt.savefig(os.path.join(path,'ownerships.png'),dpi=200,bbox_inches='tight',pad_inches=0)
plt.close()
for m in range(n_models):
plt.figure(figsize=(10,10),frameon=False)
plt.imshow(I_individuals[m])
#inds = ownership[m]
inds = point_models==m
plt.scatter(kp0[inds,0],kp0[inds,1],c='black',alpha=0.5,linewidths=(1,),s=40)
#print(kp0[inds,0])
#print(kp0[inds,1])
plt.xlim([0,self.pc_w])
plt.ylim([self.pc_h,0])
plt.xticks([],[])
plt.yticks([],[])
#plt.title('Flow model {0}'.format(m))
if m == n_models-1:
outfname = 'model_homography.png'
elif m == n_models-2:
outfname = 'model_pca.png'
else:
outfname = 'model_{0:02}.png'.format(m)
plt.savefig(os.path.join(path,outfname),dpi=200,bbox_inches='tight',pad_inches=0)
plt.close()
plt.figure(figsize=(10,10),frameon=False)
plt.imshow(res,cmap='cubehelix',vmin=0,vmax=n_models-1)
#plt.title('Segmentation')
#plt.colorbar()
plt.xticks([],[])
plt.yticks([],[])
plt.savefig(os.path.join(path,'segments.png'),dpi=200,bbox_inches='tight',pad_inches=0)
plt.close()
def debug_GC(self,unaries_warp,unaries_colors,unaries_dist,tot,res,w_y,w_x,flow_u_all=None,flow_v_all=None):
try:
from mylib import viz
except:
return
n_models = unaries_warp.shape[2]
n_coeffs = self.flow_bases_u.shape[0]
n_pixels = self.flow_bases_u.shape[1]
I_warp = np.vstack([unaries_warp[:,:,m] for m in range(n_models)])
I_col = np.vstack([unaries_colors[:,:,m] for m in range(n_models)])
I_dist = np.vstack([unaries_dist[:,:,m] for m in range(n_models)])
I_tot = np.vstack([tot[:,:,m] for m in range(n_models)])
#I_tot = I_warp + I_col + I_dist
if have_plt:
plt.figure(figsize=(10,10))
plt.subplot(1,4,1)
plt.imshow(I_warp,cmap='hot')
plt.title('Warp unaries')
plt.subplot(1,4,2)
plt.imshow(I_col,cmap='hot')
plt.title('Color model unaries')
plt.subplot(1,4,3)
plt.imshow(I_dist,cmap='hot')
plt.title('Distance unaries')
plt.subplot(1,4,4)
plt.imshow(I_tot,cmap='hot')
plt.title('Combined unaries')
plt.savefig('unaries.png',dpi=200,bbox_inches='tight')
plt.close()
if res is not None:
plt.figure(figsize=(10,10))
plt.imshow(res,cmap='hot')
plt.title('Model results')
plt.savefig('result_models.png',dpi=200,bbox_inches='tight')
plt.close()
plt.figure(figsize=(10,10))
plt.subplot(2,1,1)
plt.imshow(w_y,cmap='hot')
plt.colorbar()
plt.title('Vertical regularization weights')
plt.subplot(2,1,2)
plt.imshow(w_x,cmap='hot')
plt.colorbar()
plt.title('Horizontal reg weights')
plt.savefig('reg_weights.png',dpi=200,bbox_inches='tight')
plt.close()
plt.figure(figsize=(10,10))
plt.subplot(411)
plt.imshow(np.argmin(unaries_warp,axis=2),cmap='hot')
plt.title('min of warp unaries')
plt.subplot(412)
plt.imshow(np.argmin(unaries_colors,axis=2),cmap='hot')
plt.title('min of color unaries')
plt.subplot(413)
plt.imshow(np.argmin(unaries_dist,axis=2),cmap='hot')
plt.title('min of dist unaries')
plt.subplot(414)
plt.imshow(np.argmin(unaries_warp + unaries_colors + unaries_dist,axis=2),cmap='hot')
plt.title('min of combined unaries')
plt.savefig('unaries_min.png',dpi=200,bbox_inches='tight')
plt.close()
if flow_u_all is not None:
plt.figure(figsize=(10,10))
nx = 3
ny = int(np.ceil(float(n_models)/3.0))
for n in range(n_models):
u = flow_u_all[n].reshape((256,512))
v = flow_v_all[n].reshape((256,512))
Iv = viz.viz_flow(u,v)
plt.subplot(ny,nx,n+1)
plt.imshow(Iv)
plt.title('Model {}'.format(n))
plt.savefig('./models_all.png',dpi=200,bbox_inches='tight')
plt.close()