def create_camera_matrix():
"""
All parameters are taken from https://github.com/guiggh/hand_pose_action
"""
# camera parameter for image sensor
# Image center
u0 = 935.732544
v0 = 540.681030
# Focal Length
fx = 1395.749023
fy = 1395.749268
rgb_camera_intr = camera.CameraIntr(**{
"u0": u0,
"v0": v0,
"fx": fx,
"fy": fy
})
# camera parameter for depth sensor
# Image center:
depth_u0 = 315.944855
depth_v0 = 245.287079
# Focal Length:
depth_fx = 475.065948
depth_fy = 475.065857
depth_camera_intr = camera.CameraIntr(**{
"u0": depth_u0,
"v0": depth_v0,
"fx": depth_fx,
"fy": depth_fy
})
# Extrinsic parameters
# Rotations
R = np.array([
[0.999988496304, -0.00468848412856, 0.000982563360594],
[0.00469115935266, 0.999985218048, -0.00273845880292],
[-0.000969709653873, 0.00274303671904, 0.99999576807],
])
# Translate
t = np.array([25.7, 1.22, 3.902])
camera_extr = camera.CameraExtr(R, t)
cameras = {}
cameras["rgb_camera_intr"] = rgb_camera_intr
cameras["depth_camera_intr"] = depth_camera_intr
cameras["camera_extr"] = camera_extr
return cameras
def load_annotations(self, dataset_dir, debug=False):
logger.info("load annotations mode = {}".format(self.mode))
logger.info("dataset_dir = {}".format(dataset_dir))
annotations = []
with open(os.path.join(dataset_dir, "training_xyz.json"), 'r') as f:
joint_annotations = json.load(f)
with open(os.path.join(dataset_dir, "training_K.json"), 'r') as f:
camera_annotations = json.load(f)
global VAL_LIST
VAL_LIST = np.random.choice(len(joint_annotations),
int(0.1 * len(joint_annotations)))
for idx in range(len(joint_annotations)):
if self.mode == "train" and idx in VAL_LIST:
continue
if self.mode != "train" and idx not in VAL_LIST:
continue
rgb_joint = np.array(joint_annotations[idx])
[[fx, _, u0], [_, fy, v0], [_, _, _]] = camera_annotations[idx]
rgb_camera = camera.CameraIntr(u0=u0, v0=v0, fx=fx, fy=fy)
if DATA_CONVENTION == "ZYX":
rgb_joint = rgb_joint[:, ::-1]
rgb_joint = rgb_joint[INDEX_CONVERTER]
rgb_path = os.path.join(dataset_dir, "training", "rgb",
"{:08d}.jpg".format(idx))
rgb_joint = rgb_joint[INDEX_CONVERTER]
rgb_joint = wrist2palm(rgb_joint)
example = {}
example["hand_side"] = "right"
example["rgb_path"] = rgb_path
example["rgb_joint"] = 1000 * rgb_joint
example["rgb_camera"] = rgb_camera
annotations.append(example)
if debug and len(annotations) > 10:
break
return annotations
def create_camera_mat():
# These values are Taken from anno_all["K"]
fx = fy = 283.1
u0 = v0 = 160.
rgb_camera_intr = camera.CameraIntr(**{
"u0": u0,
"v0": v0,
"fx": fx,
"fy": fy
})
depth_camera_intr = camera.CameraIntr(**{
"u0": u0,
"v0": v0,
"fx": fx,
"fy": fy
})
cameras = {}
# Both, rgb and depth, have same camera parameter
cameras["rgb_camera_intr"] = rgb_camera_intr
cameras["depth_camera_intr"] = depth_camera_intr
return cameras
def create_camera_mat():
fx = 614.878
fy = 615.479
u0 = 313.219
v0 = 231.288
rgb_camera_intr = camera.CameraIntr(**{
"u0": u0,
"v0": v0,
"fx": fx,
"fy": fy
})
cameras = {"rgb_camera_intr": rgb_camera_intr}
return cameras
def create_camera_mat():
fx = fy = 617.173
u0 = 315.453
v0 = 242.259
rgb_camera_intr = camera.CameraIntr(**{
"u0": u0,
"v0": v0,
"fx": fx,
"fy": fy
})
fx = fy = 475.62
u0 = 311.125
v0 = 245.965
depth_camera_intr = camera.CameraIntr(**{
"u0": u0,
"v0": v0,
"fx": fx,
"fy": fy
})
cameras = {}
cameras["rgb_camera_intr"] = rgb_camera_intr
cameras["depth_camera_intr"] = depth_camera_intr
return cameras
def create_camera_mat():
# See README.txt
camera_intr = np.array([
[617.173, 0, 315.453],
[0, 617.173, 242.259],
[0, 0, 1],
])
fx = fy = 617.173
u0 = 315.453
v0 = 242.259
rgb_camera_intr = camera.CameraIntr(**{
"u0": u0,
"v0": v0,
"fx": fx,
"fy": fy
})
cameras = {}
cameras["rgb_camera_intr"] = rgb_camera_intr
return cameras
def create_camera_mat():
"""
Camera parameter of Intel Real Sense F200 active depth camera.
These parameters are taken from readme.txt of STB dataset.
1. Camera parameters
(1) Point Grey Bumblebee2 stereo camera:
base line = 120.054
fx = 822.79041
fy = 822.79041
tx = 318.47345
ty = 250.31296
(2) Intel Real Sense F200 active depth camera:
fx color = 607.92271
fy color = 607.88192
tx color = 314.78337
ty color = 236.42484
fx depth = 475.62768
fy depth = 474.77709
tx depth = 336.41179
ty depth = 238.77962
rotation vector = [0.00531 -0.01196 0.00301] (use Rodrigues' rotation formula to transform it into rotation matrix)
translation vector = [-24.0381 -0.4563 -1.2326]
(rotation and translation vector can transform the coordinates relative to color camera to those relative to depth camera)
"""
fx_color = 607.92271
fy_color = 607.88192
tx_color = 314.78337
ty_color = 236.42484
rgb_camera_intr = camera.CameraIntr(**{
"u0": tx_color,
"v0": ty_color,
"fx": fx_color,
"fy": fy_color
})
fx_depth = 475.62768
fy_depth = 474.77709
tx_depth = 336.41179
ty_depth = 238.77962
depth_camera_intr = camera.CameraIntr(**{
"u0": tx_depth,
"v0": ty_depth,
"fx": fx_depth,
"fy": fy_depth
})
def Rodrigues(rotation_vector):
theta = np.linalg.norm(rotation_vector)
rv = rotation_vector / theta
rr = np.array([[rv[i] * rv[j] for j in range(3)] for i in range(3)])
R = np.cos(theta) * np.eye(3)
R += (1 - np.cos(theta)) * rr
R += np.sin(theta) * np.array([
[0, -rv[2], rv[1]],
[rv[2], 0, -rv[0]],
[-rv[1], rv[0], 0],
])
return R
rotation_vector = np.array([0.00531, -0.01196, 0.00301])
R = Rodrigues(rotation_vector)
# transpose to fit our camera class interface
R = R.transpose()
# apply minus to fit our camera class interface
translation_vector = -np.array([-24.0381, -0.4563, -1.2326])
rgb_camera_extr = camera.CameraExtr(R, translation_vector)
depth_camera_extr = camera.CameraExtr(R, np.zeros(3))
cameras = {}
cameras["rgb_camera_intr"] = rgb_camera_intr
cameras["depth_camera_intr"] = depth_camera_intr
cameras["rgb_camera_extr"] = rgb_camera_extr
cameras["depth_camera_extr"] = depth_camera_extr
return cameras
from detector.hand_dataset.bbox_dataset import HandBBoxDataset as HandDataset
from detector.graphics import camera
from detector.hand_dataset.common_dataset import NUM_KEYPOINTS, STANDARD_KEYPOINT_NAMES, COLOR_MAP, EDGES, ROOT_IDX, REF_EDGE
from detector.hand_dataset.common_dataset import make_keypoint_converter, wrist2palm
from detector.hand_dataset.bbox_dataset import HandBBoxDataset as HandDataset
from detector.hand_dataset.geometry_utils import DATA_CONVENTION
# Puseudo image center and focal length
# This dataset provides only 2D image. to utilize our script that utilize 3D information,
# We will create puseudo 3D coordinate and camera matrix
U0 = V0 = 300
F = 500
RGB_CAMERA_INTR = camera.CameraIntr(fx=F, fy=F, u0=U0, v0=V0)
KEYPOINT_NAMES = STANDARD_KEYPOINT_NAMES
class HandDBBaseDataset(DatasetMixin):
def __init__(self, dataset_dir, mode="train", debug=False):
self.mode = mode
self.rgb_camera = RGB_CAMERA_INTR
self.depth_camera = None
self.n_joints = NUM_KEYPOINTS
self.root_idx = ROOT_IDX
self.ref_edge = REF_EDGE
self.annotations = self.load_annotations(dataset_dir, debug=debug)
def __len__(self):
return len(self.annotations)