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['obj_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
samples = self.render_images(scene_objs,
stable_pose,
self.config['images_per_stable_pose'],
camera_config=self._camera_configs())
for sample in samples:
self.save_samples(sample, grasps, T_obj_stp, collision_checker, grasp_metrics, stable_pose)
self.cur_pose_label += 1
gc.collect()
# next stable pose
self.cur_obj_label += 1
# next object
# Save dataset
self.tensor_dataset.flush()
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 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.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'))
to_frame="primesense_overhead")
# T_world_camera = RigidTransform(rotation = np.array([[1,0,0],[0,1,0],[0,0,1]]),
# translation=np.array([-.2,0,0]).T,
# from_frame="world",
# to_frame="primesense_overhead")
# squares = [(-1,-1), (-1,0), (-1,1), (0,-1), (0,1), (1,-1), (1,0), (1,1)]
# for i, square in enumerate(squares):
T_rook2_world = RigidTransform(rotation=np.eye(3),
translation=np.array([.06, .06, 0]).T,
from_frame="rook2",
to_frame="obj")
of2 = ObjFile(dexnet_path + "/data/meshes/chess_pieces/WizRook.obj")
rook2 = of2.read()
camera.add_to_scene("rook2", SceneObject(rook2, T_rook2_world))
T_table_world = RigidTransform(rotation=np.eye(3),
translation=np.array([0, 0, -.0025]).T,
from_frame="table",
to_frame="obj")
of3 = ObjFile(dexnet_path +
"/data/meshes/Table_triangulated_faces_resized.obj")
table = of3.read()
camera.add_to_scene("table", SceneObject(table, T_table_world))
of3 = ObjFile(dexnet_path + "/data/meshes/chess_pieces/WizRook.obj")
rook3 = of3.read()
T_rook3_world = RigidTransform(rotation=np.eye(3),
translation=np.array([-.06, .06, 0]).T,
from_frame="rook3",
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()
obj_key = dataset.object_keys[0]
obj = dataset[obj_key]
# read in the stable poses of the mesh
stable_poses = dataset.stable_poses(obj.key)
stable_pose = stable_poses[0]
# 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')
# sample images from random variable
T_table_obj = RigidTransform(from_frame='table', to_frame='obj')
scene_objs = {'table': SceneObject(table_mesh, T_table_obj)}
# Set up image renderer
deltas = np.arange(0, 0.4, 0.05)
camera_dist = config['env_rv_params']['min_radius']
for delta in deltas:
config['env_rv_params']['min_radius'] = camera_dist + delta
config['env_rv_params']['max_radius'] = camera_dist + delta
urv = UniformPlanarWorksurfaceImageRandomVariable(obj.mesh,
[RenderMode.DEPTH_SCENE],
'camera',
config['env_rv_params'],
scene_objs=scene_objs,
stable_pose=stable_pose)
# Render images
specular=0.25,
T_light_camera=T_light_camera,
cutoff=180)
# create material props
mat_props = MaterialProperties(color=(249, 241, 21),
ambient=0.5,
diffuse=1.0,
specular=1,
shininess=0)
# create scene objects
scene_objs = {
'table':
SceneObject(table_mesh,
T_obj_world.inverse(),
mat_props=table_mat_props)
}
for name, scene_obj in scene_objs.iteritems():
virtual_camera.add_to_scene(name, scene_obj)
# camera pose
cam_dist = 0.3
T_camera_world = RigidTransform(rotation=np.array([[0, 1,
0], [1, 0, 0],
[0, 0, -1]]),
translation=[0, 0, cam_dist],
from_frame=camera_intr.frame,
to_frame='world')
T_obj_camera = T_camera_world.inverse() * T_obj_world