def load_from_json(filename):
"""
Loads a synthetic MLA dataset
Parameters:
-----------
filename: string
Filename of the scene configuration file (format: JSON)
Returns:
--------
lenses: dictionary
Lens dictionary, keys are the axial coordinates
"""
basedir = os.path.dirname(filename)
with open(filename, 'r') as f:
lenses_json = json.load(f)
lenses = dict()
if len(lenses_json) == 0:
return lenses
diam = lenses_json[0]['diameter']
radius = diam / 2.0
inner_radius = radius - lenses_json[0]['lens_border']
local_grid = rtxlens.LocalLensGrid(diam)
# local grid coordinates shared by all lenses
x, y = local_grid.x, local_grid.y
xx, yy = local_grid.xx, local_grid.yy
mask = np.zeros_like(local_grid.xx)
mask[xx**2 + yy**2 < inner_radius**2] = 1
for lens in lenses_json:
# image data
color_filename = "{0}/{1}".format(basedir,
lens['relative_color_filename'])
# ground truth depth
depth_filename = "{0}/{1}".format(basedir,
lens['relative_depth_filename'])
tmp_lens = rtxlens.Lens(lens['axial_coord'], lens['pixel_coord'],
lens['diameter'], lens['focal_type'])
tmp_lens.col_img = plt.imread(color_filename)
if tmp_lens.col_img.shape[2] == 4:
tmp_lens.col_img = tmp_lens.col_img[:, :, :3]
weights = np.array([0.3, 0.59, 0.11])
tmp_lens.img = np.sum([
tmp_lens.col_img[:, :, i].astype(np.float64) * weights[i]
for i in range(3)
],
axis=0)
tmp_lens.grid = local_grid
tmp_lens.mask = mask
tmp_lens.num_channels = tmp_lens.col_img.shape[2]
tmp_lens.img_interp = sinterp.RectBivariateSpline(y, x, tmp_lens.img)
tmp_lens.img_interp3c[0] = sinterp.RectBivariateSpline(
y, x, tmp_lens.col_img[:, :, 0])
tmp_lens.img_interp3c[1] = sinterp.RectBivariateSpline(
y, x, tmp_lens.col_img[:, :, 1])
tmp_lens.img_interp3c[2] = sinterp.RectBivariateSpline(
y, x, tmp_lens.col_img[:, :, 2])
tmp_lens.inner_radius = inner_radius
tmp_lens.focal_length = lens['focal_length']
tmp_lens.focus_distance = lens['focus_distance']
tmp_lens.fstop = lens['fstop']
tmp_lens.pixel_size = lens['pixel_size']
tmp_lens.position = np.array(lens['position'])
#tmp_lens.position = np.array(lens['location']) # some datasets needs 'location'
tmp_lens.rotation = np.array(lens['rotation_mat']).reshape((3, 3))
#tmp_lens.rotation = np.array([1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0]).reshape((3,3)) #
tmp_lens.focus_disparity = tmp_lens.focal_length * tmp_lens.diameter / tmp_lens.focus_distance
# read the ground truth depth. The 16-Bit PNG depth files
# contain the clipped depth values, convert to the real depth
# values
tmp_depth = plt.imread(depth_filename).astype(np.float64)
tmp_depth = tmp_depth * (lens['clip_end'] -
lens['clip_start']) + lens['clip_start']
tmp_lens.depth_img = tmp_depth
tmp_lens.disp_img = np.copy(tmp_depth)
tmp_lens.disp_img[tmp_depth > 0] = (
tmp_lens.focal_length *
tmp_lens.diameter) / tmp_lens.disp_img[tmp_depth > 0]
# set the lens border
tmp_lens.lens_border = lens['lens_border']
# finally add the lens to the dictionary
lenses[tuple(lens['axial_coord'])] = tmp_lens
return lenses
def load_from_xml(image_filename, config_filename):
img = plt.imread(image_filename)
calib = read_calibration(config_filename)
if len(img.shape) < 3:
print("This version works only for colored image")
raise ValueError("Unsopported image dimension {0}".format(img.shape))
elif len(img.shape) == 3:
weights = np.array([0.3, 0.59, 0.11])
data = np.sum([img[:, :, i] * weights[i] for i in range(3)], axis=0)
data_col = img
else:
raise ValueError("Unsupported image dimensions {0}".format(img.shape))
# the image grid in pixels used for the bivariate spline interpolation
gridy, gridx = range(data.shape[0]), range(data.shape[1])
data_interp = sinterp.RectBivariateSpline(gridy, gridx, data)
data_interp_r = sinterp.RectBivariateSpline(gridy, gridx, data_col[:, :,
0])
data_interp_g = sinterp.RectBivariateSpline(gridy, gridx, data_col[:, :,
1])
data_interp_b = sinterp.RectBivariateSpline(gridy, gridx, data_col[:, :,
2])
img_shape = np.asarray(data.shape[0:2])
coords = rtxhexgrid.hex_lens_grid(img_shape, calib.lens_diameter,
calib.rot_angle, calib.offset,
calib.lbasis)
focal_type_offsets = [
tuple(lens['offset'].values()) for lens in calib.lens_types
]
focal_type_offsets = [(int(p[0]), int(p[1])) for p in focal_type_offsets]
focal_type_map = {
_hex_focal_type(p): i
for i, p in enumerate(focal_type_offsets)
}
lenses = dict()
local_grid = rtxlens.LocalLensGrid(calib.lens_diameter)
x, y = local_grid.x, local_grid.y
xx, yy = local_grid.xx, local_grid.yy
mask = np.zeros_like(local_grid.xx)
mask[xx**2 + yy**2 < calib.inner_lens_radius**2] = 1
for lc in coords:
pc = coords[lc]
lens = rtxlens.Lens(lcenter=lc,
pcenter=pc,
diameter=calib.lens_diameter)
lens.focal_type = _hex_focal_type(lc)
lens.fstop = lens.diameter
lens.pixel_size = 1.0
lens.position = np.array([pc[1], pc[0], 0])
lens.img = np.ones((local_grid.shape[0], local_grid.shape[1], 3))
cen_x = round(pc[0])
cen_y = round(pc[1])
x1 = int(cen_x + round(np.min(x)))
x2 = int(cen_x + round(np.max(x)))
y1 = int(cen_y + round(np.min(y)))
y2 = int(cen_y + round(np.max(y)))
lens.num_channels = 3
lens.img = data_interp(y + pc[0], x + pc[1])
lens.img_interp = sinterp.RectBivariateSpline(y, x, lens.img)
# colored version of the interpolated image:
lens.img_interp3c[0] = data_interp_r
lens.img_interp3c[1] = data_interp_g
lens.img_interp3c[2] = data_interp_b
lens.col_img_uint = data_col[
x1:x2 + 1, y1:y2 +
1] #np.zeros((lens.img.shape[0], lens.img.shape[1], lens.num_channels))
if lens.col_img_uint.dtype == 'uint8':
new_col_img = np.zeros((lens.col_img_uint.shape))
new_col_img = lens.col_img_uint / 255.0
lens.col_img = new_col_img
else:
lens.col_img = lens.col_img_uint
lens.mask = mask
lens.grid = local_grid
# the lens area used for matching
lens.inner_radius = calib.inner_lens_radius
# minimum and maximum disparities for this lens
focal_type = focal_type_map[_hex_focal_type(lc)]
min_depth = calib.lens_types[focal_type]['depth_range']['min']
max_depth = calib.lens_types[focal_type]['depth_range']['max']
lens.min_disp = lens.diameter / max_depth
lens.max_disp = lens.diameter / min_depth
#pdb.set_trace()
lenses[tuple(lc)] = lens
return lenses