Skip to content

Commit

Permalink
updated for topk prediction
Browse files Browse the repository at this point in the history
  • Loading branch information
Charles Qi committed Sep 25, 2015
1 parent e87077b commit ab640c7
Show file tree
Hide file tree
Showing 4 changed files with 44 additions and 40 deletions.
Binary file added demo_view/chair_image.jpg
View file
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
8 changes: 5 additions & 3 deletions demo_view/run_demo_topk.py
View file
Expand Up @@ -14,10 +14,12 @@

if __name__ == '__main__':

img_filenames = [os.path.join(BASE_DIR, 'aeroplane_image.jpg')]
class_idxs = [g_shape_names.index('aeroplane')]
#img_filenames = [os.path.join(BASE_DIR, 'aeroplane_image.jpg')]
#class_idxs = [g_shape_names.index('aeroplane')]
img_filenames = [os.path.join(BASE_DIR, 'chair_image.jpg')]
class_idxs = [g_shape_names.index('chair')]
output_result_file = os.path.join(BASE_DIR, 'est-view-topk.txt')
topk = 5
topk = 2

if not os.path.exists(output_result_file):
viewpoint_topk(img_filenames, class_idxs, topk, output_result_file)
Expand Down
13 changes: 7 additions & 6 deletions demo_view/run_visualize_3dview_topk.py
View file
Expand Up @@ -20,28 +20,29 @@
# --------------------------------
#### CHANGE BEG ####
output_result_file = os.path.join(BASE_DIR, 'est-view-topk.txt')
model_obj_file = os.path.join(g_render4cnn_root_folder, 'demo_render', 'sample_model2', 'model.obj')
original_image_file = os.path.join(BASE_DIR, 'aeroplane_image.jpg')
model_obj_file = os.path.join(g_render4cnn_root_folder, 'demo_render', 'sample_model', 'model.obj')
original_image_file = os.path.join(BASE_DIR, 'chair_image.jpg')
rendered_image_file_prefix = os.path.join(BASE_DIR, 'chair_in_estimated_view')
io_redirect = '' #' > /dev/null 2>&1'
#### CHANGE END ####
# --------------------------------


# Display estimated viewpoints that are read from result file
# suppress tilt for chairs
estimated_viewpoints = [[float(x) for x in line.rstrip().split(' ')] for line in open(output_result_file,'r')]
v = estimated_viewpoints[0]
topk = len(v)/4
print("Estimated views and confidence: ")
for k in range(topk):
a,e,t,c = v[4*k : 4*k+4]
print('rank:%d, confidence:%f, azimuth:%d, elevation:%d, tilt:%d' % (k+1, c, a, e, t))
print('rank:%d, confidence:%f, azimuth:%d, elevation:%d' % (k+1, c, a, e))

# Render images in estimated views
for k in range(topk):
a,e,t = v[4*k : 4*k+3]
a,e = v[4*k : 4*k+2]
python_cmd = 'python %s -m %s -a %s -e %s -t %s -d %s -o %s' % (os.path.join(g_render4cnn_root_folder, 'demo_render', 'render_class_view.py'),
model_obj_file, str(a), str(e), str(t), str(2.0), rendered_image_file_prefix+str(k)+'.png')
model_obj_file, str(a), str(e), str(0), str(2.0), rendered_image_file_prefix+str(k)+'.png')
print ">> Running rendering command: \n \t %s" % (python_cmd)
os.system('%s %s' % (python_cmd, io_redirect))

Expand All @@ -54,7 +55,7 @@
bbox = im2.getbbox()
im2 = im2.crop(bbox)
draw = ImageDraw.Draw(im2)
draw.text((0,0), 'rank: %d, confidence: %f\nazimuth=%f, elevation=%f, tilt=%f' % (k+1, v[4*k+3], v[4*k], v[4*k+1], v[4*k+2]), (0,255,0))
draw.text((0,0), 'rank: %d, confidence: %f\nazimuth=%f, elevation=%f' % (k+1, v[4*k+3], v[4*k], v[4*k+1]), (0,255,0))
im2s.append(im2)

im1.show()
Expand Down
63 changes: 32 additions & 31 deletions view_estimation/evaluation_helper.py
View file
Expand Up @@ -88,15 +88,16 @@ def get_top_preds(prob, num_bin=8, diff_threshold=10):

# verify
duplicate = 0
for pred, _ in preds_list:
if min(abs(local_pred - pred), N - abs(local_pred - pred)) < diff_threshold:
for i, ituple in enumerate(preds_list):
ipred, iprob = ituple
if min(abs(local_pred - ipred), N - abs(local_pred - ipred)) < diff_threshold:
if local_prob > iprob:
preds_list[i] = (local_pred, local_prob) # exchange to bigger one
duplicate = 1 # too close to existed pred
break
if duplicate == 0:
preds_list.append((local_pred, local_prob))
print 'before sorted, preds_list:\n', preds_list
preds_list = sorted(preds_list, key=lambda item:item[1], reverse=True)
print 'after sorted, preds_list:\n', preds_list
return preds_list


Expand All @@ -106,44 +107,45 @@ def get_top_preds(prob, num_bin=8, diff_threshold=10):
get topk confident viewpoint predictions from probs
@input:
probs_3dview - a length-3 list of azimuth, elevation and tilt probs (each is length-360 list)
topk (K) - integer (1000=>K>=1). only return top K confident result
topk (K) - integer (K>=1). only return top K confident result
@output:
return topk_viewpoints - a length-topk list of tuples of (<viewpoint-tuple>, <confidence>), where
<viewpoint-tuple> is a length-3 tuples of azimuth, elevation and tilt angles in degree
'''
def get_topk_viewpoints(probs_3dview, topk):
assert(topk>=1 and topk<=1000)
print 'topk:\n', topk
for n in range(10):
if n*n*n >= topk: break
assert(topk>=1)
viewpoints_list = []

# secure softmax. used to normalize an array to 0~1 with sum as 1
def my_softmax(x):
ex = np.exp(x - np.max(x))
out = ex / ex.sum()
return out

print 'n:\n', n
print 'returned by get_top_preds:\n', get_top_preds(probs_3dview[0], 16, 10)
preds_list_azimuth = get_top_preds(probs_3dview[0], 16, 10)[0:n]
print 'prob azimuth:\n', probs_3dview[0]
print 'preds_list_azimuth:\n', preds_list_azimuth
preds_list_elevation = get_top_preds(probs_3dview[1], 36, 8)[0:n]
preds_list_tilt = get_top_preds(probs_3dview[2], 72, 3)[0:n]

viewpoints_list = []
for i in range(n):
for j in range(n):
for k in range(n):
apred,aprob = preds_list_azimuth[i]
epred,eprob = preds_list_elevation[j]
tpred,tprob = preds_list_tilt[k]
viewpoints_list.append(((apred,epred,tpred),np.log(aprob)+np.log(eprob)+np.log(tprob)))
# Rank by prob-azimuth * prob-elevation * prob-tilt
#for n in range(10):
# if n*n*n >= topk: break
#probs_3dview = [my_softmax(prob) for prob in probs_3dview]
#preds_list_azimuth = get_top_preds(probs_3dview[0], 16, 45)[0:n]
#preds_list_elevation = get_top_preds(probs_3dview[1], 24, 10)[0:n]
#preds_list_tilt = get_top_preds(probs_3dview[2], 36, 5)[0:n]
#for i in range(n):
# for j in range(n):
# for k in range(n):
# apred,aprob = preds_list_azimuth[i]
# epred,eprob = preds_list_elevation[j]
# tpred,tprob = preds_list_tilt[k]
# viewpoints_list.append(((apred,epred,tpred), np.log(aprob)+np.log(eprob)+np.log(tprob)))

# Rank by azimuth and a bit of elevation
preds_list_azimuth = get_top_preds(probs_3dview[0], 16, 45)
for i in range(min(topk,len(preds_list_azimuth))):
apred,aprob = preds_list_azimuth[i]
viewpoints_list.append(((apred, probs_3dview[1].argmax(), probs_3dview[2].argmax()), aprob))


print 'viewpoints_list before sorted:\n', viewpoints_list
viewpoints_list = sorted(viewpoints_list, key=lambda item:item[1], reverse=True)
topk_viewpoints = viewpoints_list[0:topk]
print 'viewpoints_list after sorted:\n', viewpoints_list

return topk_viewpoints

Expand All @@ -155,7 +157,7 @@ def my_softmax(x):
@input:
img_filenames - list of image filenames
class_idxs - list of class index (0~11)
topk (K) - integer number (1000=>K>=1). choose top K prediction results (in default, set to 1)
topk (K) - integer number (10=>K>=1). choose top K prediction results (in default, set to 1)
output_result_file - string. output estimated viewpoints into this file
@output:
return preds - a lenght-len(img_filenames) list of length-topk list of tuples
Expand All @@ -179,7 +181,7 @@ def viewpoint_topk(img_filenames, class_idxs, topk=1, output_result_file=None):
result_keys = g_caffe_prob_keys
resize_dim = g_images_resize_dim
image_mean_file = g_image_mean_file
assert(topk>=1 and topk<=1000) # from 1^3 to 10^3
assert(topk>=1 and topk<=10) # from 1^3 to 10^3
assert(len(result_keys) == 3) #azimuth,elevation,tilt

# ** NETWORK FORWARD PASS **
Expand All @@ -203,15 +205,14 @@ def viewpoint_topk(img_filenames, class_idxs, topk=1, output_result_file=None):

# get topk viewpoints: length-topk list of lenght-3 tuples
topk_viewpoints = get_topk_viewpoints(probs_3dview, topk)
print 'topk_viewpoints in viewpoint_topk func:\n', topk_viewpoints
preds[i] = topk_viewpoints

if output_result_file is not None:
# OUTPUT: apred epred tpred
fout = open(output_result_file, 'w')
for i in range(len(img_filenames)):
topk_views = preds[i]
for k in range(topk):
for k in range(len(topk_views)):
va,ve,vt = topk_views[k][0]
confidence = topk_views[k][1]
fout.write('%d %d %d %f ' % (va,ve,vt,confidence))
Expand Down

0 comments on commit ab640c7

Please sign in to comment.