def get_statistics_from_txt(path, ground_truth, benchmark_name, height, width,
scale_factor, mrf_factor, min_val_disp,
max_val_disp, occlusion):
benchmark_path = path + 'benchmarks/'
basename = os.path.splitext(benchmark_name)[0]
# read the map to be evaluated
arr_final_disparity = np.loadtxt(benchmark_path + benchmark_name,
dtype=float)
arr_final_disparity = abs(arr_final_disparity)
w = arr_final_disparity.shape[1]
width_scale = float(w) / width
im = img.fromarray(arr_final_disparity)
im = im.resize((width, height), img.NEAREST)
arr_final_disparity = np.asarray(im, dtype=float)
arr_final_disparity = arr_final_disparity / (mrf_factor * width_scale)
arr_final_disparity[:, 0:occlusion] = 0
rms, rate = evaluate_disparity(arr_final_disparity, ground_truth)
arr_final_disparity[(arr_final_disparity == 0).nonzero()] = 'nan'
plt.plot_array('Final disparity map (' + basename + ')',
arr_final_disparity, min_val_disp, max_val_disp,
benchmark_path + 'maps/' + basename)
return rms, rate
def get_statistics_from_image(path, ground_truth, benchmark_name, height,
width, scale_factor, mrf_factor, min_val_disp,
max_val_disp, occlusion):
benchmark_path = path + 'benchmarks/'
basename = os.path.splitext(benchmark_name)[0]
# read found disparity map by one of benchmarks
arr_final_disparity_img = img.open(benchmark_path + benchmark_name)
img_w, img_h = arr_final_disparity_img.size
width_scale = float(img_w) / width
arr_final_disparity_img = arr_final_disparity_img.resize((width, height),
img.NEAREST)
arr_final_disparity = np.asarray(arr_final_disparity_img, dtype=float)
# disparity maps from MRF software need to be rescaled (see the scaling factor in the MRF output)
arr_final_disparity = arr_final_disparity / (mrf_factor * width_scale)
arr_final_disparity[:, 0:occlusion] = 0
rms, rate = evaluate_disparity(arr_final_disparity, ground_truth)
arr_final_disparity[(arr_final_disparity == 0).nonzero()] = 'nan'
plt.plot_array('Final disparity map (' + basename + ')',
arr_final_disparity, min_val_disp, max_val_disp,
benchmark_path + 'maps/' + basename)
return rms, rate
def get_statistics(path, scale_factor, disparity, min_val_disp, max_val_disp,
occlusion):
# size of disparity maps that need to be anlysed
width = disparity.shape[1]
height = disparity.shape[0]
# format of the benchmark entry:
#
# < local name of the disparity file >
# < MRF factor > WARNING maps from the MRF methods need to be rescaled (see the factor in the output)
benchmarks = (('BM.dat', 1), ('BP-M.png', 4), ('BP-S.png', 4),
('BP.dat', 1), ('CSBP.dat', 1), ('Expansion.png', 4),
('ICM.png', 4), ('phase-based.dat',
1), ('Swap.png', 4), ('TRW-S.png', 4))
# dictionary: name of the method -> RMS error, percentage of found values
statistics_output = {}
# get the ground-truth map of the pre-defined size with the proper scaling
ground_truth = get_ground_truth_map(path, scale_factor, width, height)
# evaluate disparity estimated by the current method
disparity = np.nan_to_num(disparity)
rms, rate = get_statistics_current_disparity(path, ground_truth, disparity,
min_val_disp, max_val_disp)
statistics_output['Current'] = (rms, rate)
# evaluate maps given by other techniques
for b in benchmarks:
ext = os.path.splitext(b[0])[1][1:]
mrf_factor = b[1]
rms = 0.0
rate = 0.0
if ext == 'png':
rms, rate = get_statistics_from_image(path, ground_truth, b[0],
height, width, scale_factor,
mrf_factor, min_val_disp,
max_val_disp, occlusion)
elif ext == 'dat':
rms, rate = get_statistics_from_txt(path, ground_truth, b[0],
height, width, scale_factor,
mrf_factor, min_val_disp,
max_val_disp, occlusion)
basename = os.path.splitext(b[0])[0]
statistics_output[basename] = (rms, rate)
ground_truth[(ground_truth == 0).nonzero()] = 'nan'
plt.plot_array('Ground truth', ground_truth, min_val_disp, max_val_disp,
path + 'benchmarks/' + 'maps/' + './ground-truth')
return statistics_output
def get_statistics_current_disparity(path, ground_truth, disparity, min_val_disp, max_val_disp):
benchmark_path = path + 'benchmarks/'
rms, rate = evaluate_disparity(disparity, ground_truth)
disparity[ (disparity == 0).nonzero() ] = 'nan'
plt.plot_array('Current disparity map', disparity, min_val_disp, max_val_disp, benchmark_path + 'maps/' + 'current-method')
return rms, rate
def get_statistics_current_disparity(path, ground_truth, disparity,
min_val_disp, max_val_disp):
benchmark_path = path + 'benchmarks/'
rms, rate = evaluate_disparity(disparity, ground_truth)
disparity[(disparity == 0).nonzero()] = 'nan'
plt.plot_array('Current disparity map', disparity, min_val_disp,
max_val_disp, benchmark_path + 'maps/' + 'current-method')
return rms, rate
def get_statistics(path, scale_factor, disparity, min_val_disp, max_val_disp, occlusion):
# size of disparity maps that need to be anlysed
width = disparity.shape[1]
height = disparity.shape[0]
# format of the benchmark entry:
#
# < local name of the disparity file >
# < MRF factor > WARNING maps from the MRF methods need to be rescaled (see the factor in the output)
benchmarks = ( ('BM.dat', 1), ('BP-M.png', 4), ('BP-S.png', 4), ('BP.dat', 1),
('CSBP.dat', 1), ('Expansion.png', 4), ('ICM.png', 4), ('phase-based.dat', 1),
('Swap.png', 4), ('TRW-S.png', 4)
)
# dictionary: name of the method -> RMS error, percentage of found values
statistics_output = {}
# get the ground-truth map of the pre-defined size with the proper scaling
ground_truth = get_ground_truth_map(path, scale_factor, width, height)
# evaluate disparity estimated by the current method
disparity = np.nan_to_num(disparity)
rms, rate = get_statistics_current_disparity(path, ground_truth, disparity, min_val_disp, max_val_disp)
statistics_output['Current'] = (rms,rate)
# evaluate maps given by other techniques
for b in benchmarks:
ext = os.path.splitext(b[0])[1][1:]
mrf_factor = b[1]
rms = 0.0
rate = 0.0
if ext == 'png':
rms, rate = get_statistics_from_image(path, ground_truth, b[0], height, width, scale_factor, mrf_factor, min_val_disp, max_val_disp, occlusion)
elif ext == 'dat':
rms, rate = get_statistics_from_txt(path, ground_truth, b[0], height, width, scale_factor, mrf_factor, min_val_disp, max_val_disp, occlusion)
basename = os.path.splitext(b[0])[0]
statistics_output[basename] = (rms,rate)
ground_truth[ (ground_truth == 0).nonzero() ] = 'nan'
plt.plot_array('Ground truth', ground_truth, min_val_disp, max_val_disp, path + 'benchmarks/' + 'maps/' + './ground-truth')
return statistics_output
def get_statistics_from_image(path, ground_truth, benchmark_name, height, width, scale_factor, mrf_factor, min_val_disp, max_val_disp, occlusion):
benchmark_path = path + 'benchmarks/'
basename = os.path.splitext(benchmark_name)[0]
# read found disparity map by one of benchmarks
arr_final_disparity_img = img.open(benchmark_path + benchmark_name)
img_w, img_h = arr_final_disparity_img.size
width_scale = float(img_w) / width
arr_final_disparity_img = arr_final_disparity_img.resize((width, height), img.NEAREST)
arr_final_disparity = np.asarray(arr_final_disparity_img, dtype=float)
# disparity maps from MRF software need to be rescaled (see the scaling factor in the MRF output)
arr_final_disparity = arr_final_disparity / (mrf_factor*width_scale)
arr_final_disparity[:,0:occlusion] = 0
rms, rate = evaluate_disparity(arr_final_disparity, ground_truth)
arr_final_disparity[ (arr_final_disparity == 0).nonzero() ] = 'nan'
plt.plot_array('Final disparity map (' + basename + ')', arr_final_disparity, min_val_disp, max_val_disp, benchmark_path + 'maps/' + basename)
return rms, rate
def get_statistics_from_txt(path, ground_truth, benchmark_name, height, width, scale_factor, mrf_factor, min_val_disp, max_val_disp, occlusion):
benchmark_path = path + 'benchmarks/'
basename = os.path.splitext(benchmark_name)[0]
# read the map to be evaluated
arr_final_disparity = np.loadtxt(benchmark_path + benchmark_name, dtype=float)
arr_final_disparity = abs(arr_final_disparity)
w = arr_final_disparity.shape[1]
width_scale = float(w) / width
im = img.fromarray(arr_final_disparity)
im = im.resize((width, height), img.NEAREST)
arr_final_disparity = np.asarray(im, dtype=float)
arr_final_disparity = arr_final_disparity / (mrf_factor*width_scale)
arr_final_disparity[:,0:occlusion] = 0
rms, rate = evaluate_disparity(arr_final_disparity, ground_truth)
arr_final_disparity[ (arr_final_disparity == 0).nonzero() ] = 'nan'
plt.plot_array('Final disparity map (' + basename + ')', arr_final_disparity, min_val_disp, max_val_disp, benchmark_path + 'maps/' + basename)
return rms, rate