def objective(args):
# args['ksize'] = int(args['ksize'])
costs = []
# for index in range(3):
# for index in range(10):
for index in range(194):
if index in [31, 82, 114]: # missing frames
continue
source = KittiMultiViewSource(index, test=False, n_frames=0)
full_shape = source.frame_ten[0].shape
frame_ten = [downsample(source.frame_ten[0], frame_down_factor),
downsample(source.frame_ten[1], frame_down_factor)]
frame_shape = frame_ten[0].shape
true_points = source.get_ground_truth_points(occluded=False)
# determine fovea shape
fovea_pixels = np.ceil(fovea_fraction * frame_shape[0] * (frame_shape[1] - values))
if fovea_pixels > 0:
fovea_height = np.minimum(frame_shape[0], np.ceil(fovea_pixels ** .5))
fovea_width = np.minimum(frame_shape[1], np.ceil(fovea_pixels / fovea_height))
fovea_shape = fovea_height, fovea_width
else:
fovea_shape = (0, 0)
# average disparity
try:
average_disp = source.get_average_disparity()
except IOError: # likely does not have a full 20 frames
print("Skipping index %d (lacks frames)" % index)
continue
average_disp = upsample_average_disp(
average_disp, frame_down_factor, frame_shape, values)
# run the filter
foveal = Foveal(average_disp, frame_down_factor, mem_down_factor,
fovea_shape, frame_shape, values,
iters=iters, fovea_levels=fovea_levels, **args)
disp, _ = foveal.process_frame(frame_ten)
# cost = cost_on_points(disp[:,values:], true_points, full_shape=full_shape)
cost = cost_on_points(disp[:,values:], true_points, average_disp,
full_shape=full_shape, clip=error_clip)
costs.append(cost)
error = np.mean(costs)
errors[arg_key(args)] = error
args_s = "{%s}" % ', '.join("%r: %s" % (k, args[k]) for k in sorted(space))
print("%s: %s" % (error, args_s))
return error
full_values = 128
frame_down_factor = 1
assert full_values % 2**frame_down_factor == 0
values = full_values / 2**frame_down_factor
sources = []
# for index in range(1):
# for index in range(30):
for index in range(194):
source = KittiMultiViewSource(index, test=False)
frame_ten = [downsample(source.frame_ten[0], frame_down_factor),
downsample(source.frame_ten[1], frame_down_factor)]
true_points = source.get_ground_truth_points(occluded=False)
sources.append((source, frame_ten, true_points))
r = np.log(10)
space = collections.OrderedDict([
('laplacian_ksize', pyll.scope.int(1 + hp.quniform('laplacian_ksize', 0, 20, 2))),
('laplacian_scale', hp.lognormal('laplacian_scale', 0, r)),
('data_weight', hp.lognormal('data_weight', -2*r, r)),
('data_max', hp.lognormal('data_max', 2*r, r)),
('data_exp', hp.lognormal('data_exp', 0, r)),
('disc_max', hp.lognormal('disc_max', r, r)),
('smooth', hp.lognormal('smooth', 0, r)),
('post_smooth', hp.lognormal('post_smooth', 0, r)),
])
full_values = 128
frame_down_factor = 1
assert full_values % 2**frame_down_factor == 0
values = full_values / 2**frame_down_factor
sources = []
# for index in range(1):
# for index in range(30):
for index in range(194):
source = KittiMultiViewSource(index, test=False)
frame_ten = [
downsample(source.frame_ten[0], frame_down_factor),
downsample(source.frame_ten[1], frame_down_factor)
]
true_points = source.get_ground_truth_points(occluded=False)
sources.append((source, frame_ten, true_points))
r = np.log(10)
space = collections.OrderedDict([
('laplacian_ksize',
pyll.scope.int(1 + hp.quniform('laplacian_ksize', 0, 20, 2))),
('laplacian_scale', hp.lognormal('laplacian_scale', 0, r)),
('data_weight', hp.lognormal('data_weight', -2 * r, r)),
('data_max', hp.lognormal('data_max', 2 * r, r)),
('data_exp', hp.lognormal('data_exp', 0, r)),
('disc_max', hp.lognormal('disc_max', r, r)),
('smooth', hp.lognormal('smooth', 0, r)),
('post_smooth', hp.lognormal('post_smooth', 0, r)),
])
def rationale():
"""
Figure that illustrates rationale for fovea approach, i.e. diminishing returns
with increasing runtime via further iterations at a given downfactor, but
possibly too-high cost of using lower downfactor.
"""
# n_frames = 194
n_frames = 50
frame_down_factor = 1
source = KittiMultiViewSource(0)
full_shape = source.frame_ten[0].shape
frame_shape = downsample(source.frame_ten[0], frame_down_factor).shape
params = {'data_weight': 0.16145115747533928, 'disc_max': 294.1504935618425, 'data_max': 32.024780646200725, 'laplacian_ksize': 3}
# iterations = [1, 2, 3]
iterations = [1, 2, 3, 4, 5, 7, 10, 15]
mean_coarse_times = []
mean_coarse_costs = []
mean_fine_times = []
mean_fine_costs = []
for j in range(len(iterations)):
print(str(iterations[j]) + ' iterations')
coarse_times = []
coarse_costs = []
fine_times = []
fine_costs = []
for i in range(n_frames):
print(' frame ' + str(i))
source = KittiMultiViewSource(i)
true_points = source.get_ground_truth_points(occluded=False)
frame_down_factor = 1
coarse_time, coarse_cost = _evaluate_frame(source.frame_ten, true_points, frame_shape,
full_shape, frame_down_factor+1, frame_down_factor, iterations[j], params)
fine_time, fine_cost = _evaluate_frame(source.frame_ten, true_points, frame_shape,
full_shape, frame_down_factor+0, frame_down_factor, iterations[j], params)
coarse_times.append(coarse_time)
coarse_costs.append(coarse_cost)
fine_times.append(fine_time)
fine_costs.append(fine_cost)
mean_coarse_times.append(np.mean(coarse_times))
mean_coarse_costs.append(np.mean(coarse_costs))
mean_fine_times.append(np.mean(fine_times))
mean_fine_costs.append(np.mean(fine_costs))
print(mean_coarse_times)
print(mean_coarse_costs)
print(mean_fine_times)
print(mean_fine_costs)
plt.plot(mean_coarse_times, mean_coarse_costs, color='k', marker='s', markersize=12)
plt.plot(mean_fine_times, mean_fine_costs, color='k', marker='o', markersize=12)
plt.xlabel('Runtime (s)', fontsize=18)
plt.ylabel('Mean absolute disparity error (pixels)', fontsize=18)
plt.gca().tick_params(labelsize='18')
plt.show()
def rationale():
"""
Figure that illustrates rationale for fovea approach, i.e. diminishing returns
with increasing runtime via further iterations at a given downfactor, but
possibly too-high cost of using lower downfactor.
"""
# n_frames = 194
n_frames = 50
frame_down_factor = 1
source = KittiMultiViewSource(0)
full_shape = source.frame_ten[0].shape
frame_shape = downsample(source.frame_ten[0], frame_down_factor).shape
params = {
'data_weight': 0.16145115747533928,
'disc_max': 294.1504935618425,
'data_max': 32.024780646200725,
'laplacian_ksize': 3
}
# iterations = [1, 2, 3]
iterations = [1, 2, 3, 4, 5, 7, 10, 15]
mean_coarse_times = []
mean_coarse_costs = []
mean_fine_times = []
mean_fine_costs = []
for j in range(len(iterations)):
print(str(iterations[j]) + ' iterations')
coarse_times = []
coarse_costs = []
fine_times = []
fine_costs = []
for i in range(n_frames):
print(' frame ' + str(i))
source = KittiMultiViewSource(i)
true_points = source.get_ground_truth_points(occluded=False)
frame_down_factor = 1
coarse_time, coarse_cost = _evaluate_frame(source.frame_ten,
true_points,
frame_shape, full_shape,
frame_down_factor + 1,
frame_down_factor,
iterations[j], params)
fine_time, fine_cost = _evaluate_frame(source.frame_ten,
true_points, frame_shape,
full_shape,
frame_down_factor + 0,
frame_down_factor,
iterations[j], params)
coarse_times.append(coarse_time)
coarse_costs.append(coarse_cost)
fine_times.append(fine_time)
fine_costs.append(fine_cost)
mean_coarse_times.append(np.mean(coarse_times))
mean_coarse_costs.append(np.mean(coarse_costs))
mean_fine_times.append(np.mean(fine_times))
mean_fine_costs.append(np.mean(fine_costs))
print(mean_coarse_times)
print(mean_coarse_costs)
print(mean_fine_times)
print(mean_fine_costs)
plt.plot(mean_coarse_times,
mean_coarse_costs,
color='k',
marker='s',
markersize=12)
plt.plot(mean_fine_times,
mean_fine_costs,
color='k',
marker='o',
markersize=12)
plt.xlabel('Runtime (s)', fontsize=18)
plt.ylabel('Mean absolute disparity error (pixels)', fontsize=18)
plt.gca().tick_params(labelsize='18')
plt.show()