def __init__(self, class_type):
self.class_type = class_type
self.mask_path = os.path.join(cfg.LINEMOD,'{}/mask/*.png'.format(class_type))
self.dir_path = os.path.join(cfg.LINEMOD_ORIG,'{}/data'.format(class_type))
dataset_pose_dir_path = os.path.join(cfg.DATA_DIR, 'dataset_poses')
os.system('mkdir -p {}'.format(dataset_pose_dir_path))
self.dataset_poses_path = os.path.join(dataset_pose_dir_path, '{}_poses.npy'.format(class_type))
blender_pose_dir_path = os.path.join(cfg.DATA_DIR, 'blender_poses')
os.system('mkdir -p {}'.format(blender_pose_dir_path))
self.blender_poses_path = os.path.join(blender_pose_dir_path, '{}_poses.npy'.format(class_type))
os.system('mkdir -p {}'.format(blender_pose_dir_path))
self.pose_transformer = PoseTransformer(class_type)
def get_plane_height(self):
if os.path.exists(self.plane_height_path):
plane_height = read_pickle(self.plane_height_path)
else:
plane_height = {}
if self.class_type in plane_height:
return plane_height[self.class_type]
else:
pose_transformer = PoseTransformer(self.class_type)
model = pose_transformer.get_blender_model()
height = np.min(model[:, -1])
plane_height[self.class_type] = height
save_pickle(plane_height, self.plane_height_path)
return height
class DataStatistics(object):
# world_to_camera_pose = np.array([[-1.19209304e-07, 1.00000000e+00, -2.98023188e-08, 1.19209304e-07],
# [-8.94069672e-08, 2.22044605e-16, -1.00000000e+00, 8.94069672e-08],
# [-1.00000000e+00, -8.94069672e-08, 1.19209304e-07, 1.00000000e+00]])
world_to_camera_pose = np.array([[-1.00000024e+00, -8.74227979e-08, -5.02429621e-15, 8.74227979e-08],
[5.02429621e-15, 1.34358856e-07, -1.00000012e+00, -1.34358856e-07],
[8.74227979e-08, -1.00000012e+00, 1.34358856e-07, 1.00000012e+00]])
def __init__(self, class_type):
self.class_type = class_type
self.mask_path = os.path.join(cfg.LINEMOD,'{}/mask/*.png'.format(class_type))
self.dir_path = os.path.join(cfg.LINEMOD_ORIG,'{}/data'.format(class_type))
dataset_pose_dir_path = os.path.join(cfg.DATA_DIR, 'dataset_poses')
os.system('mkdir -p {}'.format(dataset_pose_dir_path))
self.dataset_poses_path = os.path.join(dataset_pose_dir_path, '{}_poses.npy'.format(class_type))
blender_pose_dir_path = os.path.join(cfg.DATA_DIR, 'blender_poses')
os.system('mkdir -p {}'.format(blender_pose_dir_path))
self.blender_poses_path = os.path.join(blender_pose_dir_path, '{}_poses.npy'.format(class_type))
os.system('mkdir -p {}'.format(blender_pose_dir_path))
self.pose_transformer = PoseTransformer(class_type)
def get_proper_crop_size(self):
mask_paths = glob.glob(self.mask_path)
widths = []
heights = []
for mask_path in mask_paths:
mask = Image.open(mask_path).convert('1')
mask = np.array(mask).astype(np.int32)
row_col = np.argwhere(mask == 1)
min_row, max_row = np.min(row_col[:, 0]), np.max(row_col[:, 0])
min_col, max_col = np.min(row_col[:, 1]), np.max(row_col[:, 1])
width = max_col - min_col
height = max_row - min_row
widths.append(width)
heights.append(height)
widths = np.array(widths)
heights = np.array(heights)
print('min width: {}, max width: {}'.format(np.min(widths), np.max(widths)))
print('min height: {}, max height: {}'.format(np.min(heights), np.max(heights)))
def get_quat_translation(self, object_to_camera_pose):
object_to_camera_pose = np.append(object_to_camera_pose, [[0, 0, 0, 1]], axis=0)
world_to_camera_pose = np.append(self.world_to_camera_pose, [[0, 0, 0, 1]], axis=0)
object_to_world_pose = np.dot(np.linalg.inv(world_to_camera_pose), object_to_camera_pose)
quat = mat2quat(object_to_world_pose[:3, :3])
translation = object_to_world_pose[:3, 3]
return quat, translation
def get_dataset_poses(self):
if os.path.exists(self.dataset_poses_path):
poses = np.load(self.dataset_poses_path)
return poses[:, :3], poses[:, 3:]
eulers = []
translations = []
train_set = np.loadtxt(os.path.join(cfg.LINEMOD, '{}/training_range.txt'.format(self.class_type)),np.int32)
for idx in train_set:
rot_path = os.path.join(self.dir_path, 'rot{}.rot'.format(idx))
tra_path = os.path.join(self.dir_path, 'tra{}.tra'.format(idx))
pose = read_pose(rot_path, tra_path)
euler = self.pose_transformer.orig_pose_to_blender_euler(pose)
eulers.append(euler)
translations.append(pose[:, 3])
eulers = np.array(eulers)
translations = np.array(translations)
np.save(self.dataset_poses_path, np.concatenate([eulers, translations], axis=-1))
return eulers, translations
def sample_sphere(self, num_samples):
""" sample angles from the sphere
reference: https://zhuanlan.zhihu.com/p/25988652?group_id=828963677192491008
"""
flat_objects = ['037_scissors', '051_large_clamp', '052_extra_large_clamp']
if self.class_type in flat_objects:
begin_elevation = 30
else:
begin_elevation = 0
ratio = (begin_elevation + 90) / 180
num_points = int(num_samples // (1 - ratio))
phi = (np.sqrt(5) - 1.0) / 2.
azimuths = []
elevations = []
for n in range(num_points - num_samples, num_points):
z = 2. * n / num_points - 1.
azimuths.append(np.rad2deg(2 * np.pi * n * phi % (2 * np.pi)))
elevations.append(np.rad2deg(np.arcsin(z)))
return np.array(azimuths), np.array(elevations)
def sample_poses(self):
eulers, translations = self.get_dataset_poses()
num_samples = cfg.NUM_SYN
azimuths, elevations = self.sample_sphere(num_samples)
euler_sampler = stats.gaussian_kde(eulers.T)
eulers = euler_sampler.resample(num_samples).T
eulers[:, 0] = azimuths
eulers[:, 1] = elevations
translation_sampler = stats.gaussian_kde(translations.T)
translations = translation_sampler.resample(num_samples).T
np.save(self.blender_poses_path, np.concatenate([eulers, translations], axis=-1))
from base_utils import PoseTransformer, read_pose, read_pickle, save_pickle
import cv2
def fuse(img, mask, background):
background = cv2.resize(background,(img.shape[1], img.shape[0]))
silhouette = mask > 0
background[silhouette] = img[silhouette]
return background
class_type = 'cat'
dir_path = os.path.join(cfg.LINEMOD_ORIG,'{}/data'.format(class_type))
train_set = np.loadtxt(os.path.join(cfg.LINEMOD, '{}/training_range.txt'.format(class_type)),np.int32)
trans = PoseTransformer(class_type)
for idx in train_set:
rot_path = os.path.join(dir_path, 'rot{}.rot'.format(idx))
tra_path = os.path.join(dir_path, 'tra{}.tra'.format(idx))
pose = read_pose(rot_path, tra_path)
pose = trans.orig_pose_to_blender_pose(pose)
rot, tra = pose[:, :3], pose[:, 3]
break
r = OpenGLRenderer()
rgb, mask = r.render(class_type, pose, intrinsic_matrix=r.intrinsic_matrix['linemod'], render_type='all')
rgb = rgb[:,:,[2,1,0]] # opencv use bgr order instead of rgb
background = cv2.imread('Lena.png', 1)
rgb = fuse(rgb, mask, background)
cv2.imwrite('RGB2.jpg', rgb)