class Reprojection:
def __init__(self, path, elev):
if not os.path.exists(DATA_DIR):
raise NameError(
"Path %s is not specified. Is the docker file set up properly?"
% DATA_DIR)
table_file = DATA_DIR + '/meshes/table.obj'
self.table_mesh = ObjFile(table_file).read()
self.data_dir = DATA_DIR + '/meshes/dexnet/'
self.output_dir = DATA_DIR + '/reprojections/'
self.config = YamlConfig(
'./cfg/tools/generate_projected_gqcnn_dataset.yaml')
self.random_positions = 1
self.image_size = (300, 300)
self.elev = 0
self.show_images = False
self.cur_obj_label = 0
self.cur_image_label = 0
self.cur_pose_label = 0
if path is not None:
self.output_dir = DATA_DIR + '/reprojections/' + path
if elev is not None:
print("Elevation angle is being set to %d" % elev)
self.config['env_rv_params']['min_elev'] = elev
self.config['env_rv_params']['max_elev'] = elev
self.elev = elev
tensor_config = self.config['tensors']
tensor_config['fields']['depth_ims_tf_table']['height'] = 32
tensor_config['fields']['depth_ims_tf_table']['width'] = 32
tensor_config['fields']['robust_ferrari_canny'] = {}
tensor_config['fields']['robust_ferrari_canny']['dtype'] = 'float32'
tensor_config['fields']['ferrari_canny'] = {}
tensor_config['fields']['ferrari_canny']['dtype'] = 'float32'
tensor_config['fields']['force_closure'] = {}
tensor_config['fields']['force_closure']['dtype'] = 'float32'
self.tensor_dataset = TensorDataset(self.output_dir, tensor_config)
self.tensor_datapoint = self.tensor_dataset.datapoint_template
if not os.path.exists(self.output_dir + '/images'):
os.makedirs(self.output_dir + '/images')
if not os.path.exists(self.output_dir + '/meshes'):
os.makedirs(self.output_dir + '/meshes')
def _load_file_ids(self):
dtype = [('Obj_id', (np.str_, 40)), ('Tensor', int), ('Array', int),
('Depth', float)]
with open(self.data_dir + 'files.csv', 'r') as csv_file:
data = []
csv_reader = csv.reader(csv_file, delimiter=',')
for row in csv_reader:
data.append(
tuple([
int(value) if value.isdigit() else value
for value in row
]))
self.file_arr = np.array(data, dtype=dtype)
files = os.listdir(self.data_dir)
files = [name.split('.')[0] for name in files]
files.remove('files')
self.all_objects = list(set(files))
def _load_data(self, obj_id):
self.grasp_metrics = json.load(
open(self.data_dir + obj_id + '.json', 'r'))
self.candidate_grasps_dict = pkl.load(
open(self.data_dir + obj_id + '.pkl', 'rb'))
self.object_mesh = ObjFile(self.data_dir + obj_id + '.obj').read()
self.stable_poses = StablePoseFile(self.data_dir + obj_id +
'.stp').read()
self.tensor = self.file_arr['Tensor'][np.where(
self.file_arr['Obj_id'] == obj_id)][0]
self.array = self.file_arr['Array'][np.where(
self.file_arr['Obj_id'] == obj_id)][0]
depth = self.file_arr['Depth'][np.where(
self.file_arr['Obj_id'] == obj_id)][0]
self.config['env_rv_params']['min_radius'] = depth
self.config['env_rv_params']['max_radius'] = depth
def start_rendering(self):
self._load_file_ids()
for object_id in self.all_objects:
self._load_data(object_id)
for i, stable_pose in enumerate(self.stable_poses):
try:
candidate_grasp_info = self.candidate_grasps_dict[
stable_pose.id]
except KeyError:
continue
if not candidate_grasp_info:
Warning("Candidate grasp info of object id %s empty" %
object_id)
Warning("Continue.")
continue
T_obj_stp = stable_pose.T_obj_table.as_frames('obj', 'stp')
T_obj_stp = self.object_mesh.get_T_surface_obj(T_obj_stp)
T_table_obj = RigidTransform(from_frame='table',
to_frame='obj')
scene_objs = {
'table': SceneObject(self.table_mesh, T_table_obj)
}
urv = UniformPlanarWorksurfaceImageRandomVariable(
self.object_mesh,
[RenderMode.DEPTH_SCENE, RenderMode.SEGMASK],
'camera',
self.config['env_rv_params'],
scene_objs=scene_objs,
stable_pose=stable_pose)
render_sample = urv.rvs(size=self.random_positions)
# for render_sample in render_samples:
binary_im = render_sample.renders[RenderMode.SEGMASK].image
depth_im = render_sample.renders[
RenderMode.DEPTH_SCENE].image.crop(300, 300)
orig_im = Image.fromarray(self._scale_image(depth_im.data))
if self.show_images:
orig_im.show()
orig_im.convert('RGB').save(self.output_dir + '/images/' +
object_id + '_elev_' +
str(self.elev) + '_original.png')
print("Saved original")
T_stp_camera = render_sample.camera.object_to_camera_pose
shifted_camera_intr = render_sample.camera.camera_intr.crop(
300, 300, 240, 320)
depth_points = self._reproject_to_3D(depth_im,
shifted_camera_intr)
transformed_points, T_camera = self._transformation(
depth_points)
camera_dir = np.dot(T_camera.rotation,
np.array([0.0, 0.0, -1.0]))
pcd = o3d.geometry.PointCloud()
# print(camera_dir)
pcd.points = o3d.utility.Vector3dVector(transformed_points.T)
# TODO check normals!!
# pcd.estimate_normals(search_param=o3d.geometry.KDTreeSearchParamHybrid(radius=0.1, max_nn=30))
# pcd.normals = o3d.utility.Vector3dVector(-np.asarray(pcd.normals))
normals = np.repeat([camera_dir],
len(transformed_points.T),
axis=0)
pcd.normals = o3d.utility.Vector3dVector(normals)
# cs_points = [[0, 0, 0], [1, 0, 0], [0, 1, 0], [0, 0, 1]]
# cs_lines = [[0, 1], [0, 2], [0, 3]]
# colors = [[1, 0, 0], [0, 1, 0], [0, 0, 1]]
# cs = o3d.geometry.LineSet(points=o3d.utility.Vector3dVector(cs_points),
# lines=o3d.utility.Vector2iVector(cs_lines))
# cs.colors = o3d.utility.Vector3dVector(colors)
# o3d.visualization.draw_geometries([pcd])
depth = self._o3d_meshing(pcd)
# projected_depth_im,new_camera_intr,table_height = self._projection(new_points,shifted_camera_intr)
new_camera_intr = shifted_camera_intr
new_camera_intr.cx = 150
new_camera_intr.cy = 150
projected_depth_im = np.asarray(depth)
projected_depth_im[projected_depth_im == 0.0] = -1.0
table_height = np.median(
projected_depth_im[projected_depth_im != -1.0].flatten())
print("Minimum depth:", min(projected_depth_im.flatten()))
print("Maximum depth:", max(projected_depth_im.flatten()))
im = Image.fromarray(self._scale_image(projected_depth_im))
projected_depth_im = DepthImage(projected_depth_im,
frame='new_camera')
cx = projected_depth_im.center[1]
cy = projected_depth_im.center[0]
# Grasp conversion
T_obj_old_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
T_obj_camera = T_camera.dot(T_obj_old_camera)
for grasp_info in candidate_grasp_info:
grasp = grasp_info.grasp
collision_free = grasp_info.collision_free
grasp_2d = grasp.project_camera(T_obj_camera,
new_camera_intr)
dx = cx - grasp_2d.center.x
dy = cy - grasp_2d.center.y
translation = np.array([dy, dx])
# Project 3D old_camera_cs contact points into new camera cs
contact_points = np.append(grasp_info.contact_point1, 1).T
new_cam = np.dot(T_obj_camera.matrix, contact_points)
c1 = new_camera_intr.project(
Point(new_cam[0:3], frame=new_camera_intr.frame))
contact_points = np.append(grasp_info.contact_point2, 1).T
new_cam = np.dot(T_obj_camera.matrix, contact_points)
c2 = new_camera_intr.project(
Point(new_cam[0:3], frame=new_camera_intr.frame))
# Check if there are occlusions at contact points
if projected_depth_im.data[
c1.x, c1.y] == -1.0 or projected_depth_im.data[
c2.x, c2.y] == -1.0:
print("Contact point at occlusion")
contact_occlusion = True
else:
contact_occlusion = False
# Mark contact points in image
im = im.convert('RGB')
if False:
im_draw = ImageDraw.Draw(im)
im_draw.line([(c1[0], c1[1] - 10),
(c1[0], c1[1] + 10)],
fill=(255, 0, 0, 255))
im_draw.line([(c1[0] - 10, c1[1]),
(c1[0] + 10, c1[1])],
fill=(255, 0, 0, 255))
im_draw.line([(c2[0], c2[1] - 10),
(c2[0], c2[1] + 10)],
fill=(255, 0, 0, 255))
im_draw.line([(c2[0] - 10, c2[1]),
(c2[0] + 10, c2[1])],
fill=(255, 0, 0, 255))
if self.show_images:
im.show()
im.save(self.output_dir + '/images/' + object_id +
'_elev_' + str(self.elev) + '_reprojected.png')
# Transform and crop image
depth_im_tf = projected_depth_im.transform(
translation, grasp_2d.angle)
depth_im_tf = depth_im_tf.crop(96, 96)
# Apply transformation to contact points
trans_map = np.array([[1, 0, dx], [0, 1, dy]])
rot_map = cv2.getRotationMatrix2D(
(cx, cy), np.rad2deg(grasp_2d.angle), 1)
trans_map_aff = np.r_[trans_map, [[0, 0, 1]]]
rot_map_aff = np.r_[rot_map, [[0, 0, 1]]]
full_map = rot_map_aff.dot(trans_map_aff)
# print("Full map",full_map)
c1_rotated = (np.dot(full_map, np.r_[c1.vector, [1]]) -
np.array([150 - 48, 150 - 48, 0])) / 3
c2_rotated = (np.dot(full_map, np.r_[c2.vector, [1]]) -
np.array([150 - 48, 150 - 48, 0])) / 3
grasp_line = depth_im_tf.data[48]
occlusions = len(np.where(np.squeeze(grasp_line) == -1)[0])
# Set occlusions to table height for resizing image
depth_im_tf.data[depth_im_tf.data == -1.0] = table_height
depth_image = Image.fromarray(np.asarray(depth_im_tf.data))\
.resize((32, 32), resample=Image.BILINEAR)
depth_im_tf_table = np.asarray(depth_image).reshape(
32, 32, 1)
# depth_im_tf_table = depth_im_tf.resize((32, 32,), interp='bilinear')
im = Image.fromarray(
self._scale_image(depth_im_tf_table.reshape(
32, 32))).convert('RGB')
draw = ImageDraw.Draw(im)
draw.line([(c1_rotated[0], c1_rotated[1] - 3),
(c1_rotated[0], c1_rotated[1] + 3)],
fill=(255, 0, 0, 255))
draw.line([(c1_rotated[0] - 3, c1_rotated[1]),
(c1_rotated[0] + 3, c1_rotated[1])],
fill=(255, 0, 0, 255))
draw.line([(c2_rotated[0], c2_rotated[1] - 3),
(c2_rotated[0], c2_rotated[1] + 3)],
fill=(255, 0, 0, 255))
draw.line([(c2_rotated[0] - 3, c2_rotated[1]),
(c2_rotated[0] + 3, c2_rotated[1])],
fill=(255, 0, 0, 255))
if self.show_images:
im.show()
im.save(self.output_dir + '/images/' + object_id +
'_elev_' + str(self.elev) + '_transformed.png')
hand_pose = np.r_[grasp_2d.center.y, grasp_2d.center.x,
grasp_2d.depth, grasp_2d.angle,
grasp_2d.center.y - new_camera_intr.cy,
grasp_2d.center.x - new_camera_intr.cx,
grasp_2d.width_px / 3]
self.tensor_datapoint[
'depth_ims_tf_table'] = depth_im_tf_table
self.tensor_datapoint['hand_poses'] = hand_pose
self.tensor_datapoint['obj_labels'] = self.cur_obj_label
self.tensor_datapoint['collision_free'] = collision_free
self.tensor_datapoint['pose_labels'] = self.cur_pose_label
self.tensor_datapoint[
'image_labels'] = self.cur_image_label
self.tensor_datapoint['files'] = [self.tensor, self.array]
self.tensor_datapoint['occlusions'] = occlusions
self.tensor_datapoint[
'contact_occlusion'] = contact_occlusion
for metric_name, metric_val in self.grasp_metrics[str(
grasp.id)].iteritems():
coll_free_metric = (1 * collision_free) * metric_val
self.tensor_datapoint[metric_name] = coll_free_metric
self.tensor_dataset.add(self.tensor_datapoint)
print("Saved dataset point")
self.cur_image_label += 1
self.cur_pose_label += 1
gc.collect()
self.cur_obj_label += 1
self.tensor_dataset.flush()
def _o3d_meshing(self, pcd):
mesh, densities = o3d.geometry.TriangleMesh.create_from_point_cloud_poisson(
pcd, depth=15)
densities = np.asarray(densities)
# print('visualize densities')
# densities = np.asarray(densities)
# density_colors = plt.get_cmap('plasma')(
# (densities - densities.min()) / (densities.max() - densities.min()))
# density_colors = density_colors[:, :3]
# density_mesh = o3d.geometry.TriangleMesh()
# density_mesh.vertices = mesh.vertices
# density_mesh.triangles = mesh.triangles
# density_mesh.triangle_normals = mesh.triangle_normals
# density_mesh.vertex_colors = o3d.utility.Vector3dVector(density_colors)
# o3d.visualization.draw_geometries([density_mesh])
vertices_to_remove = densities < 7.0 # np.quantile(densities, 0.01)
mesh.remove_vertices_by_mask(vertices_to_remove)
mesh.compute_vertex_normals()
mesh.paint_uniform_color([0.6, 0.6, 0.6])
o3d.io.write_triangle_mesh(
self.output_dir + '/meshes/' + "%05d_%03d.ply" %
(self.tensor, self.array), mesh)
if visualise_mesh:
o3d.visualization.draw_geometries([mesh])
vis = o3d.visualization.Visualizer()
vis.create_window(height=300, width=300, visible=False)
vis.get_render_option().load_from_json("./data/renderconfig.json")
vis.add_geometry(mesh)
vic = vis.get_view_control()
params = vic.convert_to_pinhole_camera_parameters()
(fx, fy) = params.intrinsic.get_focal_length()
(cx, cy) = params.intrinsic.get_principal_point()
params.intrinsic.set_intrinsics(300, 300, 525, 525, cx, cy)
params.extrinsic = np.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0],
[0, 0, 0, 1]])
vic.convert_from_pinhole_camera_parameters(params)
vis.poll_events()
vis.update_renderer()
depth = vis.capture_depth_float_buffer(do_render=True)
# vis.destroy_window()
# del vis
return depth
def _projection(self, transformed_points, camera_intr):
# Use Camera intrinsics
new_camera_intr = camera_intr
new_camera_intr.cx = 150
new_camera_intr.cy = 150
K = new_camera_intr.proj_matrix
projected_points = np.dot(K, transformed_points)
point_depths = projected_points[2, :]
table = np.median(point_depths)
point_z = np.tile(point_depths, [3, 1])
points_proj = np.divide(projected_points, point_z)
# Rounding
points_proj = np.round(points_proj)
points_proj = points_proj[:2, :].astype(np.int16)
valid_ind = np.where((points_proj[0, :] >= 0)
& (points_proj[0, :] < self.image_size[0])
& (points_proj[1, :] >= 0)
& (points_proj[1, :] < self.image_size[0]))[0]
# Fill new image with NaN's
fill_value = -1.0
depth_data = np.full([self.image_size[0], self.image_size[1]],
fill_value)
for ind in valid_ind:
prev_depth = depth_data[points_proj[1, ind], points_proj[0, ind]]
if prev_depth == fill_value or prev_depth >= point_depths[ind]:
depth_data[points_proj[1, ind],
points_proj[0, ind]] = point_depths[ind]
return depth_data, new_camera_intr, table
def _transformation(self, points):
# Points are given in camera frame. Transform to new camera frame!
camera_new_position = np.array([[0], [0], [0]])
ang = np.deg2rad(self.elev)
Rot = np.array([[1, 0, 0], [0, np.cos(ang), -np.sin(ang)],
[0, np.sin(ang), np.cos(ang)]])
dist = self.config['env_rv_params']['min_radius'] * np.sin(ang)
height = self.config['env_rv_params']['min_radius'] - self.config[
'env_rv_params']['min_radius'] * np.cos(ang)
# Rotation around x axis, therefore translation back to object center alongside y axis
trans = np.array([0, dist, height])
Rt = np.column_stack((Rot, trans))
# Apply transformation to homogeneous coordinates of the points
homogeneous_points = np.append(np.transpose(points),
np.ones((1, len(points))),
axis=0)
transformed_points = np.dot(Rt, homogeneous_points)
return transformed_points, RigidTransform(rotation=Rot,
translation=trans,
from_frame='camera',
to_frame='new_camera')
# def _PCA(self, points, sorting=True):
# mean = np.mean(points, axis=0)
# data_adjusted = points - mean
#
# matrix = np.cov(data_adjusted.T)
# eigenvalues, eigenvectors = np.linalg.eig(matrix)
#
# if sorting:
# sort = eigenvalues.argsort()[::-1]
# eigenvalues = eigenvalues[sort]
# eigenvectors = eigenvectors[:, sort]
# return eigenvalues, eigenvectors
def _reproject_to_3D(self, depth_im, camera_intr):
# depth points will be given in camera frame!
depth_points = camera_intr.deproject(depth_im).data.T
return depth_points
def _scale_image(self, depth):
size = depth.shape
flattend = depth.flatten()
scaled = np.interp(flattend, (0.5, 0.75), (0, 255))
integ = scaled.astype(np.uint8)
integ.resize(size)
return integ
def generate_gqcnn_dataset(dataset_path, database, target_object_keys,
env_rv_params, gripper_name, config):
"""
Generates a GQ-CNN TensorDataset for training models with new grippers, quality metrics, objects, and cameras.
Parameters
----------
dataset_path : str
path to save the dataset to
database : :obj:`Hdf5Database`
Dex-Net database containing the 3D meshes, grasps, and grasp metrics
target_object_keys : :obj:`OrderedDict`
dictionary mapping dataset names to target objects (either 'all' or a list of specific object keys)
env_rv_params : :obj:`OrderedDict`
parameters of the camera and object random variables used in sampling (see meshpy_berkeley.UniformPlanarWorksurfaceImageRandomVariable for more info)
gripper_name : str
name of the gripper to use
config : :obj:`autolab_core.YamlConfig`
other parameters for dataset generation
Notes
-----
Required parameters of config are specified in Other Parameters
Other Parameters
----------------
images_per_stable_pose : int
number of object and camera poses to sample for each stable pose
stable_pose_min_p : float
minimum probability of occurrence for a stable pose to be used in data generation (used to prune bad stable poses
gqcnn/crop_width : int
width, in pixels, of crop region around each grasp center, before resize (changes the size of the region seen by the GQ-CNN)
gqcnn/crop_height : int
height, in pixels, of crop region around each grasp center, before resize (changes the size of the region seen by the GQ-CNN)
gqcnn/final_width : int
width, in pixels, of final transformed grasp image for input to the GQ-CNN (defaults to 32)
gqcnn/final_height : int
height, in pixels, of final transformed grasp image for input to the GQ-CNN (defaults to 32)
table_alignment/max_approach_table_angle : float
max angle between the grasp axis and the table normal when the grasp approach is maximally aligned with the table normal
table_alignment/max_approach_offset : float
max deviation from perpendicular approach direction to use in grasp collision checking
table_alignment/num_approach_offset_samples : int
number of approach samples to use in collision checking
collision_checking/table_offset : float
max allowable interpenetration between the gripper and table to be considered collision free
collision_checking/table_mesh_filename : str
path to a table mesh for collision checking (default data/meshes/table.obj)
collision_checking/approach_dist : float
distance, in meters, between the approach pose and final grasp pose along the grasp axis
collision_checking/delta_approach : float
amount, in meters, to discretize the straight-line path from the gripper approach pose to the final grasp pose
tensors/datapoints_per_file : int
number of datapoints to store in each unique tensor file on disk
tensors/fields : :obj:`dict`
dictionary mapping field names to dictionaries specifying the data type, height, width, and number of channels for each tensor
debug : bool
True (or 1) if the random seed should be set to enforce deterministic behavior, False (0) otherwise
vis/candidate_grasps : bool
True (or 1) if the collision free candidate grasps should be displayed in 3D (for debugging)
vis/rendered_images : bool
True (or 1) if the rendered images for each stable pose should be displayed (for debugging)
vis/grasp_images : bool
True (or 1) if the transformed grasp images should be displayed (for debugging)
"""
# read data gen params
output_dir = dataset_path
gripper = RobotGripper.load(gripper_name)
image_samples_per_stable_pose = config['images_per_stable_pose']
stable_pose_min_p = config['stable_pose_min_p']
# read gqcnn params
gqcnn_params = config['gqcnn']
im_crop_height = gqcnn_params['crop_height']
im_crop_width = gqcnn_params['crop_width']
im_final_height = gqcnn_params['final_height']
im_final_width = gqcnn_params['final_width']
cx_crop = float(im_crop_width) / 2
cy_crop = float(im_crop_height) / 2
# open database
dataset_names = target_object_keys.keys()
datasets = [database.dataset(dn) for dn in dataset_names]
# set target objects
for dataset in datasets:
if target_object_keys[dataset.name] == 'all':
target_object_keys[dataset.name] = dataset.object_keys
# setup grasp params
table_alignment_params = config['table_alignment']
min_grasp_approach_offset = -np.deg2rad(
table_alignment_params['max_approach_offset'])
max_grasp_approach_offset = np.deg2rad(
table_alignment_params['max_approach_offset'])
max_grasp_approach_table_angle = np.deg2rad(
table_alignment_params['max_approach_table_angle'])
num_grasp_approach_samples = table_alignment_params[
'num_approach_offset_samples']
phi_offsets = []
if max_grasp_approach_offset == min_grasp_approach_offset:
phi_inc = 1
elif num_grasp_approach_samples == 1:
phi_inc = max_grasp_approach_offset - min_grasp_approach_offset + 1
else:
phi_inc = (max_grasp_approach_offset - min_grasp_approach_offset) / (
num_grasp_approach_samples - 1)
phi = min_grasp_approach_offset
while phi <= max_grasp_approach_offset:
phi_offsets.append(phi)
phi += phi_inc
# setup collision checking
coll_check_params = config['collision_checking']
approach_dist = coll_check_params['approach_dist']
delta_approach = coll_check_params['delta_approach']
table_offset = coll_check_params['table_offset']
table_mesh_filename = coll_check_params['table_mesh_filename']
if not os.path.isabs(table_mesh_filename):
table_mesh_filename = os.path.join(
os.path.dirname(os.path.realpath(__file__)), '..',
table_mesh_filename)
table_mesh = ObjFile(table_mesh_filename).read()
# set tensor dataset config
tensor_config = config['tensors']
tensor_config['fields']['depth_ims_tf_table']['height'] = im_final_height
tensor_config['fields']['depth_ims_tf_table']['width'] = im_final_width
tensor_config['fields']['obj_masks']['height'] = im_final_height
tensor_config['fields']['obj_masks']['width'] = im_final_width
# add available metrics (assuming same are computed for all objects)
metric_names = []
dataset = datasets[0]
obj_keys = dataset.object_keys
if len(obj_keys) == 0:
raise ValueError('No valid objects in dataset %s' % (dataset.name))
obj = dataset[obj_keys[0]]
grasps = dataset.grasps(obj.key, gripper=gripper.name)
grasp_metrics = dataset.grasp_metrics(obj.key,
grasps,
gripper=gripper.name)
metric_names = grasp_metrics[grasp_metrics.keys()[0]].keys()
for metric_name in metric_names:
tensor_config['fields'][metric_name] = {}
tensor_config['fields'][metric_name]['dtype'] = 'float32'
# init tensor dataset
tensor_dataset = TensorDataset(output_dir, tensor_config)
tensor_datapoint = tensor_dataset.datapoint_template
# setup log file
experiment_log_filename = os.path.join(output_dir,
'dataset_generation.log')
formatter = logging.Formatter('%(asctime)s %(levelname)s: %(message)s')
hdlr = logging.FileHandler(experiment_log_filename)
hdlr.setFormatter(formatter)
logging.getLogger().addHandler(hdlr)
root_logger = logging.getLogger()
# copy config
out_config_filename = os.path.join(output_dir, 'dataset_generation.json')
ordered_dict_config = collections.OrderedDict()
for key in config.keys():
ordered_dict_config[key] = config[key]
with open(out_config_filename, 'w') as outfile:
json.dump(ordered_dict_config, outfile)
# 1. Precompute the set of valid grasps for each stable pose:
# i) Perpendicular to the table
# ii) Collision-free along the approach direction
# load grasps if they already exist
grasp_cache_filename = os.path.join(output_dir, CACHE_FILENAME)
if os.path.exists(grasp_cache_filename):
logging.info('Loading grasp candidates from file')
candidate_grasps_dict = pkl.load(open(grasp_cache_filename, 'rb'))
# otherwise re-compute by reading from the database and enforcing constraints
else:
# create grasps dict
candidate_grasps_dict = {}
# loop through datasets and objects
for dataset in datasets:
logging.info('Reading dataset %s' % (dataset.name))
for obj in dataset:
if obj.key not in target_object_keys[dataset.name]:
continue
# init candidate grasp storage
candidate_grasps_dict[obj.key] = {}
# setup collision checker
collision_checker = GraspCollisionChecker(gripper)
collision_checker.set_graspable_object(obj)
# read in the stable poses of the mesh
stable_poses = dataset.stable_poses(obj.key)
for i, stable_pose in enumerate(stable_poses):
# render images if stable pose is valid
if stable_pose.p > stable_pose_min_p:
candidate_grasps_dict[obj.key][stable_pose.id] = []
# setup table in collision checker
T_obj_stp = stable_pose.T_obj_table.as_frames(
'obj', 'stp')
T_obj_table = obj.mesh.get_T_surface_obj(
T_obj_stp,
delta=table_offset).as_frames('obj', 'table')
T_table_obj = T_obj_table.inverse()
collision_checker.set_table(table_mesh_filename,
T_table_obj)
# read grasp and metrics
grasps = dataset.grasps(obj.key, gripper=gripper.name)
logging.info(
'Aligning %d grasps for object %s in stable %s' %
(len(grasps), obj.key, stable_pose.id))
# align grasps with the table
aligned_grasps = [
grasp.perpendicular_table(stable_pose)
for grasp in grasps
]
# check grasp validity
logging.info(
'Checking collisions for %d grasps for object %s in stable %s'
% (len(grasps), obj.key, stable_pose.id))
for aligned_grasp in aligned_grasps:
# check angle with table plane and skip unaligned grasps
_, grasp_approach_table_angle, _ = aligned_grasp.grasp_angles_from_stp_z(
stable_pose)
perpendicular_table = (
np.abs(grasp_approach_table_angle) <
max_grasp_approach_table_angle)
if not perpendicular_table:
continue
# check whether any valid approach directions are collision free
collision_free = False
for phi_offset in phi_offsets:
rotated_grasp = aligned_grasp.grasp_y_axis_offset(
phi_offset)
collides = collision_checker.collides_along_approach(
rotated_grasp, approach_dist,
delta_approach)
if not collides:
collision_free = True
break
# store if aligned to table
candidate_grasps_dict[obj.key][
stable_pose.id].append(
GraspInfo(aligned_grasp, collision_free))
# visualize if specified
if collision_free and config['vis'][
'candidate_grasps']:
logging.info('Grasp %d' % (aligned_grasp.id))
vis.figure()
vis.gripper_on_object(gripper, aligned_grasp,
obj,
stable_pose.T_obj_world)
vis.show()
# save to file
logging.info('Saving to file')
pkl.dump(candidate_grasps_dict, open(grasp_cache_filename, 'wb'))
# 2. Render a dataset of images and associate the gripper pose with image coordinates for each grasp in the Dex-Net database
# setup variables
obj_category_map = {}
pose_category_map = {}
cur_pose_label = 0
cur_obj_label = 0
cur_image_label = 0
# render images for each stable pose of each object in the dataset
render_modes = [RenderMode.SEGMASK, RenderMode.DEPTH_SCENE]
for dataset in datasets:
logging.info('Generating data for dataset %s' % (dataset.name))
# iterate through all object keys
object_keys = dataset.object_keys
for obj_key in object_keys:
obj = dataset[obj_key]
if obj.key not in target_object_keys[dataset.name]:
continue
# read in the stable poses of the mesh
stable_poses = dataset.stable_poses(obj.key)
for i, stable_pose in enumerate(stable_poses):
# render images if stable pose is valid
if stable_pose.p > stable_pose_min_p:
# log progress
logging.info('Rendering images for object %s in %s' %
(obj.key, stable_pose.id))
# add to category maps
if obj.key not in obj_category_map.keys():
obj_category_map[obj.key] = cur_obj_label
pose_category_map['%s_%s' %
(obj.key,
stable_pose.id)] = cur_pose_label
# read in candidate grasps and metrics
candidate_grasp_info = candidate_grasps_dict[obj.key][
stable_pose.id]
candidate_grasps = [g.grasp for g in candidate_grasp_info]
grasp_metrics = dataset.grasp_metrics(obj.key,
candidate_grasps,
gripper=gripper.name)
# compute object pose relative to the table
T_obj_stp = stable_pose.T_obj_table.as_frames('obj', 'stp')
T_obj_stp = obj.mesh.get_T_surface_obj(T_obj_stp)
# sample images from random variable
T_table_obj = RigidTransform(from_frame='table',
to_frame='obj')
scene_objs = {
'table': SceneObject(table_mesh, T_table_obj)
}
urv = UniformPlanarWorksurfaceImageRandomVariable(
obj.mesh,
render_modes,
'camera',
env_rv_params,
stable_pose=stable_pose,
scene_objs=scene_objs)
render_start = time.time()
render_samples = urv.rvs(
size=image_samples_per_stable_pose)
render_stop = time.time()
logging.info('Rendering images took %.3f sec' %
(render_stop - render_start))
# visualize
if config['vis']['rendered_images']:
d = int(np.ceil(
np.sqrt(image_samples_per_stable_pose)))
# binary
vis2d.figure()
for j, render_sample in enumerate(render_samples):
vis2d.subplot(d, d, j + 1)
vis2d.imshow(render_sample.renders[
RenderMode.SEGMASK].image)
# depth table
vis2d.figure()
for j, render_sample in enumerate(render_samples):
vis2d.subplot(d, d, j + 1)
vis2d.imshow(render_sample.renders[
RenderMode.DEPTH_SCENE].image)
vis2d.show()
# tally total amount of data
num_grasps = len(candidate_grasps)
num_images = image_samples_per_stable_pose
num_save = num_images * num_grasps
logging.info('Saving %d datapoints' % (num_save))
# for each candidate grasp on the object compute the projection
# of the grasp into image space
for render_sample in render_samples:
# read images
binary_im = render_sample.renders[
RenderMode.SEGMASK].image
depth_im_table = render_sample.renders[
RenderMode.DEPTH_SCENE].image
# read camera params
T_stp_camera = render_sample.camera.object_to_camera_pose
shifted_camera_intr = render_sample.camera.camera_intr
# read pixel offsets
cx = depth_im_table.center[1]
cy = depth_im_table.center[0]
# compute intrinsics for virtual camera of the final
# cropped and rescaled images
camera_intr_scale = float(im_final_height) / float(
im_crop_height)
cropped_camera_intr = shifted_camera_intr.crop(
im_crop_height, im_crop_width, cy, cx)
final_camera_intr = cropped_camera_intr.resize(
camera_intr_scale)
# create a thumbnail for each grasp
for grasp_info in candidate_grasp_info:
# read info
grasp = grasp_info.grasp
collision_free = grasp_info.collision_free
# get the gripper pose
T_obj_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
grasp_2d = grasp.project_camera(
T_obj_camera, shifted_camera_intr)
# center images on the grasp, rotate to image x axis
dx = cx - grasp_2d.center.x
dy = cy - grasp_2d.center.y
translation = np.array([dy, dx])
binary_im_tf = binary_im.transform(
translation, grasp_2d.angle)
depth_im_tf_table = depth_im_table.transform(
translation, grasp_2d.angle)
# crop to image size
binary_im_tf = binary_im_tf.crop(
im_crop_height, im_crop_width)
depth_im_tf_table = depth_im_tf_table.crop(
im_crop_height, im_crop_width)
# resize to image size
binary_im_tf = binary_im_tf.resize(
(im_final_height, im_final_width),
interp='nearest')
depth_im_tf_table = depth_im_tf_table.resize(
(im_final_height, im_final_width))
# visualize the transformed images
if config['vis']['grasp_images']:
grasp_center = Point(
depth_im_tf_table.center,
frame=final_camera_intr.frame)
tf_grasp_2d = Grasp2D(
grasp_center,
0,
grasp_2d.depth,
width=gripper.max_width,
camera_intr=final_camera_intr)
# plot 2D grasp image
vis2d.figure()
vis2d.subplot(2, 2, 1)
vis2d.imshow(binary_im)
vis2d.grasp(grasp_2d)
vis2d.subplot(2, 2, 2)
vis2d.imshow(depth_im_table)
vis2d.grasp(grasp_2d)
vis2d.subplot(2, 2, 3)
vis2d.imshow(binary_im_tf)
vis2d.grasp(tf_grasp_2d)
vis2d.subplot(2, 2, 4)
vis2d.imshow(depth_im_tf_table)
vis2d.grasp(tf_grasp_2d)
vis2d.title('Coll Free? %d' %
(grasp_info.collision_free))
vis2d.show()
# plot 3D visualization
vis.figure()
T_obj_world = vis.mesh_stable_pose(
obj.mesh,
stable_pose.T_obj_world,
style='surface',
dim=0.5)
vis.gripper(gripper,
grasp,
T_obj_world,
color=(0.3, 0.3, 0.3))
vis.show()
# form hand pose array
hand_pose = np.r_[grasp_2d.center.y,
grasp_2d.center.x,
grasp_2d.depth, grasp_2d.angle,
grasp_2d.center.y -
shifted_camera_intr.cy,
grasp_2d.center.x -
shifted_camera_intr.cx,
grasp_2d.width_px]
# store to data buffers
tensor_datapoint[
'depth_ims_tf_table'] = depth_im_tf_table.raw_data
tensor_datapoint[
'obj_masks'] = binary_im_tf.raw_data
tensor_datapoint['hand_poses'] = hand_pose
tensor_datapoint['collision_free'] = collision_free
tensor_datapoint['obj_labels'] = cur_obj_label
tensor_datapoint['pose_labels'] = cur_pose_label
tensor_datapoint['image_labels'] = cur_image_label
for metric_name, metric_val in grasp_metrics[
grasp.id].iteritems():
coll_free_metric = (
1 * collision_free) * metric_val
tensor_datapoint[
metric_name] = coll_free_metric
tensor_dataset.add(tensor_datapoint)
# update image label
cur_image_label += 1
# update pose label
cur_pose_label += 1
# force clean up
gc.collect()
# update object label
cur_obj_label += 1
# force clean up
gc.collect()
# save last file
tensor_dataset.flush()
# save category mappings
obj_cat_filename = os.path.join(output_dir, 'object_category_map.json')
json.dump(obj_category_map, open(obj_cat_filename, 'w'))
pose_cat_filename = os.path.join(output_dir, 'pose_category_map.json')
json.dump(pose_category_map, open(pose_cat_filename, 'w'))
class GenerateBalancedObliqueDataset:
def __init__(self, output_dir):
self.NUM_OBJECTS = None
self.table_file = DATA_DIR + '/meshes/table.obj'
self.data_dir = DATA_DIR + '/meshes/dexnet/'
if not os.path.exists(output_dir):
os.mkdir(output_dir)
self.image_dir = output_dir + '/images/'
if not os.path.exists(self.image_dir):
os.mkdir(self.image_dir)
self.orig_image_dir = output_dir + '/orig_images/'
if not os.path.exists(self.orig_image_dir):
os.mkdir(self.orig_image_dir)
self.config = YamlConfig(CONFIG)
self.phi_offsets = self._generate_phi_offsets()
self.datasets = self._load_datasets()
self.tensor_dataset = TensorDataset(output_dir, self._set_tensor_config())
self.tensor_datapoint = self.tensor_dataset.datapoint_template
self.gripper = self._set_gripper()
self._table_mesh_filename = self._set_table_mesh_filename()
self.table_mesh = self._set_table_mesh()
self.cur_pose_label = 0
self.cur_obj_label = 0
self.cur_image_label = 0
self.cur_image_file = 0
self.obj = None
self.T_obj_camera = None
def _camera_configs(self):
return self.config['env_rv_params'].copy()
@property
def _camera_intr_scale(self):
return 32.0 / 96.0
@property
def _render_modes(self):
return [RenderMode.DEPTH_SCENE]
@property
def _max_grasp_approach_offset(self):
return np.deg2rad(self.config['table_alignment']['max_approach_offset'])
@property
def _min_grasp_approach_offset(self):
return -np.deg2rad(self.config['table_alignment']['max_approach_offset'])
@property
def _max_grasp_approach_table_angle(self):
return np.deg2rad(self.config['table_alignment']['max_approach_table_angle'])
@property
def _num_grasp_approach_samples(self):
return self.config['table_alignment']['num_approach_offset_samples']
def _generate_phi_offsets(self):
phi_offsets = []
if self._max_grasp_approach_offset == self._min_grasp_approach_offset:
phi_inc = 1
elif self._num_grasp_approach_samples == 1:
phi_inc = self._max_grasp_approach_offset - self._min_grasp_approach_offset + 1
else:
phi_inc = (self._max_grasp_approach_offset - self._min_grasp_approach_offset) / \
(self._num_grasp_approach_samples - 1)
phi = self._min_grasp_approach_offset
while phi <= self._max_grasp_approach_offset:
phi_offsets.append(phi)
phi += phi_inc
return phi_offsets
@property
def _approach_dist(self):
return self.config['collision_checking']['approach_dist']
@property
def _delta_approach(self):
return self.config['collision_checking']['delta_approach']
@property
def _table_offset(self):
return self.config['collision_checking']['table_offset']
@property
def _stable_pose_min_p(self):
return self.config['stable_pose_min_p']
def _set_table_mesh_filename(self):
table_mesh_filename = self.config['collision_checking']['table_mesh_filename']
if not os.path.isabs(table_mesh_filename):
return os.path.join(DATA_DIR, table_mesh_filename)
return table_mesh_filename
def _set_table_mesh(self):
return ObjFile(self._table_mesh_filename).read()
def _set_gripper(self):
return RobotGripper.load(self.config['gripper'])
def _load_datasets(self):
""" Load the datasets in the Hdf5 database given through the config file.
Returns
-------
datasets (list): list of datasets
"""
database = Hdf5Database(self.config['database_name'], access_level=READ_ONLY_ACCESS)
target_object_keys = self.config['target_objects']
dataset_names = target_object_keys.keys()
datasets = [database.dataset(dn) for dn in dataset_names]
if self.NUM_OBJECTS is not None:
datasets = [dataset.subset(0, self.NUM_OBJECTS) for dataset in datasets]
return datasets
def _set_tensor_config(self):
""" Sets the tensor config based on the used config file.
Returns
-------
tensor_config (dict): tensor config that can be used to initiate a tensor dataset.
"""
tensor_config = self.config['tensors']
tensor_config['fields']['robust_ferrari_canny'] = {}
tensor_config['fields']['robust_ferrari_canny']['dtype'] = 'float32'
return tensor_config
def render_images(self, scene_objs, stable_pose, num_images, camera_config=None):
""" Renders depth and binary images from self.obj at the given stable pose. The camera
sampling occurs within urv.rvs.
Parameters
----------
scene_objs (dict): Objects occuring in the scene, mostly includes the table mesh.
stable_pose (StablePose): Stable pose of the object
num_images (int): Numbers of images to render
camera_config (dict): Camera sampling parameters with minimum/maximum values for radius, polar angle, ...
Returns
-------
render_samples (list): list of rendered images including sampled camera positions
"""
if camera_config is None:
camera_config = self.config['env_rv_params']
urv = UniformPlanarWorksurfaceImageRandomVariable(self.obj.mesh,
self._render_modes,
'camera',
camera_config,
scene_objs=scene_objs,
stable_pose=stable_pose)
# Render images
render_samples = urv.rvs(size=num_images)
return render_samples
def align_grasps_with_camera(self, grasps):
""" Attempts to align all grasps in grasps with the z axis of the camera in self.T_obj_camera.
Parameters
----------
grasps (list): List of grasps for the object mesh in the object frame.
Returns
-------
aligned_grasps (list): List of aligned grasps for the object mesh in the object frame
"""
z_axis_in_obj = np.dot(self.T_obj_camera.inverse().matrix, np.array((0, 0, -1, 1)).reshape(4, 1))
z_axis = z_axis_in_obj[0:3].squeeze() / np.linalg.norm(z_axis_in_obj[0:3].squeeze())
aligned_grasps = [grasp.perpendicular_table(z_axis) for grasp in grasps]
return aligned_grasps
def get_camera_pose(self, sample):
""" Attempts to align all grasps in grasps with the z axis of the camera in self.T_obj_camera.
Parameters
----------
sample (meshpy rendersample): Image sample
Returns
-------
camera pose (np.array): Array including the camera radius, elevation angle, polar angle, roll, focal length,
and table translation in x and y.
"""
return np.r_[sample.camera.radius,
sample.camera.elev,
sample.camera.az,
sample.camera.roll,
sample.camera.focal,
sample.camera.tx,
sample.camera.ty]
def is_grasp_aligned(self, aligned_grasp, stable_pose=None):
""" Checks if the grasp is aligned with the camera z axis or the z axis of the stable pose, if given.
Parameters
----------
aligned_grasp (Grasp)
stable_pose (StablePose): stable pose of object. Default: None
Returns
-------
Aligned (bool): True if grasp is aligned with the desired z axis.
"""
if stable_pose is not None:
_, grasp_approach_table_angle, _ = aligned_grasp.grasp_angles_from_stp_z(stable_pose)
else:
_, grasp_approach_table_angle, _ = aligned_grasp.grasp_angles_from_camera_z(self.T_obj_camera)
perpendicular_table = (np.abs(grasp_approach_table_angle) < self._max_grasp_approach_table_angle)
if not perpendicular_table:
return False
return True
def is_grasp_collision_free(self, grasp, collision_checker):
""" Checks if any of the +- phi grasp linear approaches are collision free with OpenRave collision checker.
Parameters
----------
grasp (Grasp)
collision_checker (GraspCollisionChecker)
Returns
-------
collision_free (bool): True if grasp is collision free.
"""
collision_free = False
for phi_offset in self.phi_offsets:
grasp.grasp_y_axis_offset(phi_offset)
collides = collision_checker.collides_along_approach(grasp, self._approach_dist,
self._delta_approach)
if not collides:
collision_free = True
break
return collision_free
def get_hand_pose(self, grasp_center_x, grasp_center_y, grasp_2d, grasp_3d, grasp_width, grasp_width_px):
""" Organises numpy array for hand_pose tensor.
Parameters
----------
grasp_2d (Grasp2D)
grasp_3d (Grasp3D)
Returns
-------
hand_pose (np.array): Hand_pose tensor including distance between camera and grasp tcp and grasp rotation
quaternion in camera frame.
"""
return np.r_[grasp_center_x,
grasp_center_y,
grasp_2d.depth,
grasp_3d.quaternion, # w x y z layout
grasp_width,
grasp_width_px]
def _crop_and_resize(self, image):
""" Crop and resize an image to the final height and width.
Parameters
----------
image (DepthImage)
Returns
-------
final_im (np.array): cropped and resized according to crop_height/width and final_height/width in
self.config['gqcnn']
"""
cropped_im = image.crop(self.config['gqcnn']['crop_height'], self.config['gqcnn']['crop_width'])
resized_im = Image.fromarray(np.asarray(cropped_im.data)).resize((self.config['gqcnn']['final_height'],
self.config['gqcnn']['final_width']),
resample=Image.BILINEAR)
final_im = np.asarray(resized_im).reshape(self.config['gqcnn']['final_height'],
self.config['gqcnn']['final_width'],
1)
return final_im
@staticmethod
def scale(array):
""" Scale image for visualisation purposes."""
size = array.shape
flattend = array.flatten()
scaled = np.interp(flattend, (min(flattend), max(flattend)), (0, 255), left=0, right=255)
integ = scaled.astype(np.uint8)
integ.resize(size)
return integ.squeeze()
def _show_image(self, image):
""" Show image by saving it in /data/test_image.png"""
scaled_image = self.scale(image)
im = Image.fromarray(scaled_image).resize((300, 300), resample=Image.NEAREST)
im.save('/data/test_image.png')
im.show()
def prepare_images(self, depth_im_table, dx, dy):
# Get translation image distances to grasp
translation = np.array([dy, dx])
# Transform, crop and resize image
depth_im_tf_table = depth_im_table.transform(translation, 0.0)
depth_im_tf = self._crop_and_resize(depth_im_tf_table)
# self._show_image(depth_im_tf)
# self._show_image(depth_unrotated_im_tf)
return depth_im_tf
def add_datapoint(self, hand_pose,
collision_free, aligned_grasp, grasp_metrics):
# Add data to tensor dataset
self.tensor_datapoint['hand_poses'] = hand_pose
self.tensor_datapoint['collision_free'] = collision_free
# Add metrics to tensor dataset
self.tensor_datapoint['robust_ferrari_canny'] = (1 * collision_free) * \
grasp_metrics[aligned_grasp.id]['robust_ferrari_canny']
self.tensor_dataset.add(self.tensor_datapoint)
def width_px(self, camera_intr, point1, point2):
"""Returns the width in pixels."""
# Form the jaw locations in 3D space at the given depth.
p1 = Point(point1, frame='camera')
p2 = Point(point2, frame='camera')
# Project into pixel space.
u1 = camera_intr.project(p1)
u2 = camera_intr.project(p2)
return np.linalg.norm(u1.data - u2.data)
def save_samples(self, sample, grasps, T_obj_stp, collision_checker, grasp_metrics, stable_pose):
self.tensor_datapoint['image_labels'] = self.cur_image_label
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames('obj', T_stp_camera.from_frame)
aligned_grasps = self.align_grasps_with_camera(grasps)
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
cx = depth_im_table.center[1]
cy = depth_im_table.center[0]
camera_pose = self.get_camera_pose(sample)
self.tensor_datapoint['camera_poses'] = camera_pose
shifted_camera_intr = sample.camera.camera_intr
self.tensor_datapoint['camera_intrs'] = np.r_[shifted_camera_intr.fx,
shifted_camera_intr.fy,
shifted_camera_intr.cx,
shifted_camera_intr.cy]
self.save_orig_image(depth_im_table, camera_pose, stable_pose, shifted_camera_intr)
grid_space = 200
grids = 3
centre = np.empty((grids, grids, 2))
saved = np.ones((grids, grids)) * -1
depth_images = np.empty((grids, grids, 32, 32, 1))
# Crop images in a grid around the centre
for x_grid in range(grids):
x_pos = grid_space / 4 - x_grid * grid_space / (grids + 1)
for y_grid in range(grids):
y_pos = grid_space / 4 - y_grid * grid_space / (grids + 1)
depth_im = self.prepare_images(depth_im_table, x_pos, y_pos)
self._show_image(depth_im)
depth_images[x_grid, y_grid, :, :, :] = depth_im
centre[x_grid, y_grid, 0] = x_pos
centre[x_grid, y_grid, 1] = y_pos
for cnt, aligned_grasp in enumerate(aligned_grasps):
collision_free = self.is_grasp_collision_free(aligned_grasp, collision_checker)
# Project grasp coordinates in image
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera, shifted_camera_intr)
grasp_point = aligned_grasp.close_fingers(self.obj, vis=False, check_approach=False)
if not grasp_point[0]:
Warning("Could not find contact points")
continue
# Get grasp width in pixel from endpoints
p_1 = np.dot(self.T_obj_camera.matrix, np.append(grasp_point[1][0].point, 1).T).T[:3]
p_2 = np.dot(self.T_obj_camera.matrix, np.append(grasp_point[1][1].point, 1).T).T[:3]
grasp_width_px = self.width_px(shifted_camera_intr, p_1, p_2)
grasp_width = aligned_grasp.width_from_endpoints(grasp_point[1][0].point, grasp_point[1][1].point)
pos_x = cx - grasp_2d.center.x
pos_y = cy - grasp_2d.center.y
distance = np.abs((centre[:, :, 0] - pos_x)**2 + (centre[:, :, 1] - pos_y)**2)
idx = np.where(distance == np.amin(distance))
idx_x = idx[0][0]
idx_y = idx[1][0]
grasp_center_x = (300 - grasp_2d.center.x - centre[idx_x, idx_y, 0])//3
grasp_center_y = (300 - grasp_2d.center.y - centre[idx_x, idx_y, 1])//3
if saved[idx_x, idx_y] == -1:
np.save(self.image_dir + "/depth_im_{:07d}.npy".format(self.cur_image_file),
depth_images[idx_x, idx_y, :, :, :])
saved[idx_x, idx_y] = self.cur_image_file
self.tensor_datapoint['image_files'] = self.cur_image_file
self.cur_image_file += 1
else:
self.tensor_datapoint['image_files'] = saved[idx_x, idx_y]
T_grasp_camera = aligned_grasp.gripper_pose(self.gripper).inverse() * self.T_obj_camera.inverse()
if VISUALISE_3D:
z_axis = T_grasp_camera.z_axis
theta = np.rad2deg(np.arccos((z_axis[2]) /
(np.sqrt((z_axis[0] ** 2 + z_axis[1] ** 2 + z_axis[2] ** 2)))))
vis.figure()
T_obj_world = vis.mesh_stable_pose(self.obj.mesh.trimesh,
stable_pose.T_obj_world, style='surface', dim=0.5, plot_table=True)
T_camera_world = T_obj_world * self.T_obj_camera.inverse()
vis.gripper(self.gripper, aligned_grasp, T_obj_world, color=(0.3, 0.3, 0.3),
T_camera_world=T_camera_world)
print(theta)
vis.show()
hand_pose = self.get_hand_pose(grasp_center_x, grasp_center_y, grasp_2d, T_grasp_camera,
grasp_width, grasp_width_px)
self.add_datapoint(hand_pose,
collision_free, aligned_grasp, grasp_metrics)
self.cur_image_label += 1
def save_orig_image(self, depth, camera_pose, stable_pose, camera_intr):
depth_im = Image.fromarray(depth.data).crop((225, 225, 375, 375))
depth_im.save(self.orig_image_dir + 'depth_{:06d}.tiff'.format(self.cur_image_label))
with open(self.orig_image_dir + 'object_{:06d}.txt'.format(self.cur_image_label), 'w') as f:
f.write(self.obj.key)
self.T_obj_camera.save(self.orig_image_dir + 'camera_tf_{:06d}.tf'.format(self.cur_image_label))
np.savetxt(self.orig_image_dir + 'camera_pose_{:06d}.txt'.format(self.cur_image_label), camera_pose)
save_stp = StablePoseFile(self.orig_image_dir + 'stable_pose_{:06d}.stp'.format(self.cur_image_label))
save_stp.write([stable_pose])
camera_intr.save(self.orig_image_dir + 'camera_intr_{:06d}.intr'.format(self.cur_image_label))
def render_data(self):
logging.basicConfig(level=logging.WARNING)
for dataset in self.datasets:
logging.info('Generating data for dataset %s' % dataset.name)
object_keys = dataset.object_keys
for obj_key in object_keys:
self.tensor_datapoint['object_labels'] = self.cur_obj_label
if self.cur_obj_label % 10 == 0:
logging.info("Object number: %d" % self.cur_obj_label)
self.obj = dataset[obj_key]
grasps = dataset.grasps(self.obj.key, gripper=self.gripper.name)
# Load grasp metrics
grasp_metrics = dataset.grasp_metrics(self.obj.key,
grasps,
gripper=self.gripper.name)
# setup collision checker
collision_checker = GraspCollisionChecker(self.gripper)
collision_checker.set_graspable_object(self.obj)
# read in the stable poses of the mesh
stable_poses = dataset.stable_poses(self.obj.key)
# Iterate through stable poses
for i, stable_pose in enumerate(stable_poses):
if not stable_pose.p > self._stable_pose_min_p:
continue
self.tensor_datapoint['pose_labels'] = self.cur_pose_label
# setup table in collision checker
T_obj_stp = stable_pose.T_obj_table.as_frames('obj', 'stp')
T_obj_table = self.obj.mesh.get_T_surface_obj(T_obj_stp,
delta=self._table_offset).as_frames('obj', 'table')
T_table_obj = T_obj_table.inverse()
T_obj_stp = self.obj.mesh.get_T_surface_obj(T_obj_stp)
collision_checker.set_table(self._table_mesh_filename, T_table_obj)
# sample images from random variable
T_table_obj = RigidTransform(from_frame='table', to_frame='obj')
scene_objs = {'table': SceneObject(self.table_mesh, T_table_obj)}
# Set up image renderer
before = time.time()
samples = self.render_images(scene_objs,
stable_pose,
self.config['images_per_stable_pose'],
camera_config=self._camera_configs())
# print("Rendering took {:05f} seconds".format(time.time()-before))
# times = []
for sample in samples:
# before = time.time()
self.save_samples(sample, grasps, T_obj_stp, collision_checker, grasp_metrics, stable_pose)
# times.append(time.time() - before)
# print("Saving one sample took avg {:05f} seconds".format(sum(times)/len(times)))
self.cur_pose_label += 1
gc.collect()
# next stable pose
self.cur_obj_label += 1
# next object
# Save dataset
self.tensor_dataset.flush()
# plt.subplot(132)
# plt.title('QUALITY', fontsize=FONT_SIZE)
# quality_image = quality_image[70:130,70:130,:]
# plt.imshow(quality_image[:,:,0])
# plt.subplot(133)
# plt.title('ANGLE', fontsize=FONT_SIZE)
# grasp_angle_image = grasp_angle_image[70:130,70:130,:]
# plt.imshow(grasp_angle_image[:,:,0])
# plt.show()
# break
gc.collect()
# break
tensor_dataset.flush()
#########################################################################
# np.random.shuffle(image_tensor_inds)
# for ind in image_tensor_inds:
# print 'Loading image tensor', ind
# # load image data
# image_arr = np.load(image_tensors[ind])['arr_0']
# start_ind_arr = np.load(start_ind_tensors[ind])['arr_0']
# end_ind_arr = np.load(end_ind_tensors[ind])['arr_0']
# i = np.random.choice(DATAPOINTS_PER_FILE, size=1)[0]
# print 'Showing image', i
# # read datapoint
class GenerateBalancedObliqueDataset:
def __init__(self, output_dir):
self.NUM_OBJECTS = None
self.table_file = DATA_DIR + '/meshes/table.obj'
self.data_dir = DATA_DIR + '/meshes/dexnet/'
output_dir = DATA_DIR + output_dir
if not os.path.exists(output_dir):
os.mkdir(output_dir)
self.output_dir = output_dir
self.config = YamlConfig(
'./cfg/tools/generate_oblique_gqcnn_dataset.yaml')
self.phi_offsets = self._generate_phi_offsets()
self.datasets, self.target_object_keys = self._load_datasets()
self.tensor_dataset = TensorDataset(self.output_dir,
self._set_tensor_config())
self.tensor_datapoint = self.tensor_dataset.datapoint_template
self.gripper = self._set_gripper()
self._table_mesh_filename = self._set_table_mesh_filename()
self.table_mesh = self._set_table_mesh()
self.cur_pose_label = 0
self.cur_obj_label = 0
self.cur_image_label = 0
self.grasp_upsample_distribuitions = self._get_upsample_distributions()
self.camera_distributions = self._get_camera_distributions()
self.obj = None
self.T_obj_camera = None
def _get_camera_distributions(self):
# Load relative amount of grasps per elevation angle phi
relative_elevation_angles = np.load(
DATA_DIR + '/meshes/relative_elevation_angles.npy')
relative_elevation_angles[relative_elevation_angles > 1] = 1.0
# Load positivity rate per elevation angle phi
positivity_rate = np.load(DATA_DIR + '/meshes/positivity_rate.npy')
# Get additional sample ratio through resampling positive grasps per elevation angle phi
additional_sample_ratio = 1 + self.config[
'gt_positive_ratio'] - positivity_rate / 100
# Add additonal sample ratio to relative amount of grasps per elevation angle phi
camera_frequencies = 1 / (relative_elevation_angles *
additional_sample_ratio)
camera_frequencies = np.round(
camera_frequencies / sum(camera_frequencies) * 10000).astype(int)
# Get camera distribution according to goal sampling ratios
camera_distribution = [0] * camera_frequencies[0] + [1] * camera_frequencies[1] + \
[2] * camera_frequencies[2] + [3] * camera_frequencies[3] + \
[4] * camera_frequencies[4] + [5] * camera_frequencies[5] + \
[6] * camera_frequencies[6] + [7] * camera_frequencies[7] + \
[8] * camera_frequencies[8] + [9] * camera_frequencies[9] + \
[10] * camera_frequencies[10] + [11] * camera_frequencies[11]
return camera_distribution
def _get_upsample_distributions(self):
# Get ratio of ground truth positive grasps in unbalanced dataset
positivity_rate = np.load(DATA_DIR + '/meshes/positivity_rate.npy')
# Get desired ratio of ground truth positive grasps
gt_positive_ratio = self.config['gt_positive_ratio'] * 100
# Get upsample ratio for positive grasps
upsample_ratio = (gt_positive_ratio -
positivity_rate) / positivity_rate
# Set upsample ratio of negative sampling ratios to zero
upsample_ratio[upsample_ratio < 0] = 0
X = [0, 1, 2, 3, 4, 5]
sample_distributions = []
# Fit poisson distribution to ratios
for ratio in upsample_ratio:
poisson_pd = np.round(poisson.pmf(X, ratio) * 1000).astype(int)
# Generate vector with possible sample sizes to randomly choose from
sample_distribution = [0] * poisson_pd[0] + [1] * poisson_pd[1] + [2] * poisson_pd[2] + \
[3] * poisson_pd[3] + [4] * poisson_pd[4] + [5] * poisson_pd[5]
sample_distributions.append(sample_distribution)
return sample_distributions
def _camera_configs(self):
return self.config['env_rv_params'].copy()
@property
def _camera_intr_scale(self):
return 32.0 / 96.0
@property
def _render_modes(self):
return [RenderMode.SEGMASK, RenderMode.DEPTH_SCENE]
@property
def _max_grasp_approach_offset(self):
return np.deg2rad(
self.config['table_alignment']['max_approach_offset'])
@property
def _min_grasp_approach_offset(self):
return -np.deg2rad(
self.config['table_alignment']['max_approach_offset'])
@property
def _max_grasp_approach_table_angle(self):
return np.deg2rad(
self.config['table_alignment']['max_approach_table_angle'])
@property
def _num_grasp_approach_samples(self):
return self.config['table_alignment']['num_approach_offset_samples']
def _generate_phi_offsets(self):
phi_offsets = []
if self._max_grasp_approach_offset == self._min_grasp_approach_offset:
phi_inc = 1
elif self._num_grasp_approach_samples == 1:
phi_inc = self._max_grasp_approach_offset - self._min_grasp_approach_offset + 1
else:
phi_inc = (self._max_grasp_approach_offset - self._min_grasp_approach_offset) / \
(self._num_grasp_approach_samples - 1)
phi = self._min_grasp_approach_offset
while phi <= self._max_grasp_approach_offset:
phi_offsets.append(phi)
phi += phi_inc
return phi_offsets
@property
def _approach_dist(self):
return self.config['collision_checking']['approach_dist']
@property
def _delta_approach(self):
return self.config['collision_checking']['delta_approach']
@property
def _table_offset(self):
return self.config['collision_checking']['table_offset']
@property
def _stable_pose_min_p(self):
return self.config['stable_pose_min_p']
def _set_table_mesh_filename(self):
table_mesh_filename = self.config['collision_checking'][
'table_mesh_filename']
if not os.path.isabs(table_mesh_filename):
return os.path.join(DATA_DIR, table_mesh_filename)
return table_mesh_filename
def _set_table_mesh(self):
return ObjFile(self._table_mesh_filename).read()
def _set_gripper(self):
return RobotGripper.load(self.config['gripper'])
def _load_datasets(self):
database = Hdf5Database(self.config['database_name'],
access_level=READ_ONLY_ACCESS)
target_object_keys = self.config['target_objects']
dataset_names = target_object_keys.keys()
datasets = [database.dataset(dn) for dn in dataset_names]
if self.NUM_OBJECTS is not None:
datasets = [
dataset.subset(0, self.NUM_OBJECTS) for dataset in datasets
]
return datasets, target_object_keys
def _set_tensor_config(self):
tensor_config = self.config['tensors']
tensor_config['fields']['depth_ims_tf_table']['height'] = 32
tensor_config['fields']['depth_ims_tf_table']['width'] = 32
tensor_config['fields']['obj_masks']['height'] = 32
tensor_config['fields']['obj_masks']['width'] = 32
tensor_config['fields']['robust_ferrari_canny'] = {}
tensor_config['fields']['robust_ferrari_canny']['dtype'] = 'float32'
tensor_config['fields']['ferrari_canny'] = {}
tensor_config['fields']['ferrari_canny']['dtype'] = 'float32'
tensor_config['fields']['force_closure'] = {}
tensor_config['fields']['force_closure']['dtype'] = 'float32'
return tensor_config
def get_positive_grasps(self, dataset, grasps):
metrics = dataset.grasp_metrics(self.obj.key,
grasps,
gripper=self.gripper.name)
positive_grasps = []
for cnt in range(0, len(metrics)):
if metrics[cnt]['robust_ferrari_canny'] >= 0.002:
positive_grasps.append(grasps[cnt])
return positive_grasps
def render_images(self,
scene_objs,
stable_pose,
num_images,
camera_config=None):
if camera_config is None:
camera_config = self.config['env_rv_params']
urv = UniformPlanarWorksurfaceImageRandomVariable(
self.obj.mesh,
self._render_modes,
'camera',
camera_config,
scene_objs=scene_objs,
stable_pose=stable_pose)
# Render images
render_samples = urv.rvs(size=num_images)
return render_samples
def align_grasps(self, grasps):
z_axis_in_obj = np.dot(self.T_obj_camera.inverse().matrix,
np.array((0, 0, -1, 1)).reshape(4, 1))
z_axis = z_axis_in_obj[0:3].squeeze() / np.linalg.norm(
z_axis_in_obj[0:3].squeeze())
aligned_grasps = [
grasp.perpendicular_table(z_axis) for grasp in grasps
]
return aligned_grasps
def get_camera_pose(self, sample):
return np.r_[sample.camera.radius, sample.camera.elev,
sample.camera.az, sample.camera.roll, sample.camera.focal,
sample.camera.tx, sample.camera.ty]
def is_grasp_aligned(self, aligned_grasp):
_, grasp_approach_table_angle, _ = aligned_grasp.grasp_angles_from_camera_z(
self.T_obj_camera)
perpendicular_table = (np.abs(grasp_approach_table_angle) <
self._max_grasp_approach_table_angle)
if not perpendicular_table:
return False
return True
def is_grasp_collision_free(self, grasp, collision_checker):
collision_free = False
for phi_offset in self.phi_offsets:
grasp.grasp_y_axis_offset(phi_offset)
collides = collision_checker.collides_along_approach(
grasp, self._approach_dist, self._delta_approach)
if not collides:
collision_free = True
break
return collision_free
def get_hand_pose(self, grasp_2d, shifted_camera_intr):
# Configure hand pose
return np.r_[grasp_2d.center.y, grasp_2d.center.x, grasp_2d.depth,
grasp_2d.angle,
grasp_2d.center.y - shifted_camera_intr.cy,
grasp_2d.center.x - shifted_camera_intr.cx,
grasp_2d.width_px / 3]
def prepare_images(self, depth_im_table, binary_im, grasp_2d, cx, cy):
# Get translation image distances to grasp
dx = cx - grasp_2d.center.x
dy = cy - grasp_2d.center.y
translation = np.array([dy, dx])
# Transform, crop and resize image
binary_im_tf = binary_im.transform(translation, grasp_2d.angle)
depth_im_tf_table = depth_im_table.transform(translation,
grasp_2d.angle)
binary_im_tf = binary_im_tf.crop(96, 96)
depth_im_tf_table = depth_im_tf_table.crop(96, 96)
depth_image = np.asarray(depth_im_tf_table.data)
dep_image = Image.fromarray(depth_image).resize(
(32, 32), resample=Image.BILINEAR)
depth_im_tf = np.asarray(dep_image).reshape(32, 32, 1)
binary_image = np.asarray(binary_im_tf.data)
bin_image = Image.fromarray(binary_image).resize(
(32, 32), resample=Image.BILINEAR)
binary_im_tf = np.asarray(bin_image).reshape(32, 32, 1)
return depth_im_tf, binary_im_tf
def add_datapoint(self, depth_im_tf, binary_im_tf, hand_pose,
collision_free, camera_pose, aligned_grasp,
grasp_metrics):
# Add data to tensor dataset
self.tensor_datapoint['depth_ims_tf_table'] = depth_im_tf
self.tensor_datapoint['obj_masks'] = binary_im_tf
self.tensor_datapoint['hand_poses'] = hand_pose
self.tensor_datapoint['obj_labels'] = self.cur_obj_label
self.tensor_datapoint['collision_free'] = collision_free
self.tensor_datapoint['pose_labels'] = self.cur_pose_label
self.tensor_datapoint['image_labels'] = self.cur_image_label
self.tensor_datapoint['camera_poses'] = camera_pose
# Add metrics to tensor dataset
for metric_name, metric_val in grasp_metrics[
aligned_grasp.id].iteritems():
coll_free_metric = (1 * collision_free) * metric_val
self.tensor_datapoint[metric_name] = coll_free_metric
self.tensor_dataset.add(self.tensor_datapoint)
def save_samples(self,
sample,
grasps,
T_obj_stp,
collision_checker,
grasp_metrics,
only_positive=False):
T_stp_camera = sample.camera.object_to_camera_pose
self.T_obj_camera = T_stp_camera * T_obj_stp.as_frames(
'obj', T_stp_camera.from_frame)
aligned_grasps = self.align_grasps(grasps)
binary_im = sample.renders[RenderMode.SEGMASK].image
depth_im_table = sample.renders[RenderMode.DEPTH_SCENE].image
camera_pose = self.get_camera_pose(sample)
shifted_camera_intr = sample.camera.camera_intr
cx = depth_im_table.center[1]
cy = depth_im_table.center[0]
for cnt, aligned_grasp in enumerate(aligned_grasps):
if not self.is_grasp_aligned(aligned_grasp):
continue
collision_free = self.is_grasp_collision_free(
aligned_grasp, collision_checker)
if only_positive and not collision_free:
continue
# Project grasp coordinates in image
grasp_2d = aligned_grasp.project_camera(self.T_obj_camera,
shifted_camera_intr)
depth_im_tf, binary_im_tf = self.prepare_images(
depth_im_table, binary_im, grasp_2d, cx, cy)
hand_pose = self.get_hand_pose(grasp_2d, shifted_camera_intr)
self.add_datapoint(depth_im_tf, binary_im_tf, hand_pose,
collision_free, camera_pose, aligned_grasp,
grasp_metrics)
self.cur_image_label += 1
def render_data(self):
logging.basicConfig(level=logging.WARNING)
for dataset in self.datasets:
logging.info('Generating data for dataset %s' % dataset.name)
object_keys = dataset.object_keys
for obj_key in object_keys:
logging.info("Object number: %d" % self.cur_obj_label)
self.obj = dataset[obj_key]
grasps = dataset.grasps(self.obj.key,
gripper=self.gripper.name)
positive_grasps = self.get_positive_grasps(dataset, grasps)
# Load grasp metrics
grasp_metrics = dataset.grasp_metrics(
self.obj.key, grasps, gripper=self.gripper.name)
# setup collision checker
collision_checker = GraspCollisionChecker(self.gripper)
collision_checker.set_graspable_object(self.obj)
# read in the stable poses of the mesh
stable_poses = dataset.stable_poses(self.obj.key)
# Iterate through stable poses
for i, stable_pose in enumerate(stable_poses):
if not stable_pose.p > self._stable_pose_min_p:
continue
# setup table in collision checker
T_obj_stp = stable_pose.T_obj_table.as_frames('obj', 'stp')
T_obj_table = self.obj.mesh.get_T_surface_obj(
T_obj_stp,
delta=self._table_offset).as_frames('obj', 'table')
T_table_obj = T_obj_table.inverse()
T_obj_stp = self.obj.mesh.get_T_surface_obj(T_obj_stp)
collision_checker.set_table(self._table_mesh_filename,
T_table_obj)
# sample images from random variable
T_table_obj = RigidTransform(from_frame='table',
to_frame='obj')
scene_objs = {
'table': SceneObject(self.table_mesh, T_table_obj)
}
# Set up image renderer
for _ in range(self.config['images_per_stable_pose']):
elev_angle = np.random.choice(
self.camera_distributions)
camera_config = self._camera_configs()
camera_config['min_elev'] = elev_angle * 5.0
camera_config['max_elev'] = (elev_angle + 1) * 5.0
sample = self.render_images(
scene_objs,
stable_pose,
1,
camera_config=camera_config)
self.save_samples(sample, grasps, T_obj_stp,
collision_checker, grasp_metrics)
# Get camera elevation angle for current sample
elev_angle = sample.camera.elev * 180 / np.pi
elev_bar = int(elev_angle // 5)
# Sample number of positive images from distribution
number_positive_images = np.random.choice(
self.grasp_upsample_distribuitions[elev_bar])
# Render only positive images
new_config = self._camera_configs()
new_config['min_elev'] = elev_angle
new_config['max_elev'] = elev_angle
positive_render_samples = self.render_images(
scene_objs,
stable_pose,
number_positive_images,
camera_config=new_config)
if type(positive_render_samples) == list:
for pos_sample in positive_render_samples:
self.save_samples(pos_sample,
positive_grasps,
T_obj_stp,
collision_checker,
grasp_metrics,
only_positive=True)
else:
self.save_samples(positive_render_samples,
positive_grasps,
T_obj_stp,
collision_checker,
grasp_metrics,
only_positive=True)
self.cur_pose_label += 1
gc.collect()
# next stable pose
self.cur_obj_label += 1
# next object
# Save dataset
self.tensor_dataset.flush()