def __getitem__(self, index):
img_path = os.path.join(self.data_root, self.ims[index])
img = imageio.imread(img_path).astype(np.float32) / 255.
msk = self.get_mask(index, img)
img = cv2.resize(img, (1000, 1002), interpolation=cv2.INTER_AREA)
msk = cv2.resize(msk, (1000, 1002), interpolation=cv2.INTER_NEAREST)
cam_ind = self.cam_inds[index]
K = np.array(self.cams['K'][cam_ind])
D = np.array(self.cams['D'][cam_ind])
img = cv2.undistort(img, K, D)
msk = cv2.undistort(msk, K, D)
R = np.array(self.cams['R'][cam_ind])
T = np.array(self.cams['T'][cam_ind]) / 1000.
# reduce the image resolution by ratio
H, W = int(img.shape[0] * cfg.ratio), int(img.shape[1] * cfg.ratio)
img = cv2.resize(img, (W, H), interpolation=cv2.INTER_AREA)
msk = cv2.resize(msk, (W, H), interpolation=cv2.INTER_NEAREST)
if cfg.mask_bkgd:
img[msk == 0] = 0
if cfg.white_bkgd:
img[msk == 0] = 1
K[:2] = K[:2] * cfg.ratio
i = int(os.path.basename(img_path)[:-4])
feature, coord, out_sh, can_bounds, bounds, Rh, Th = self.prepare_input(
i)
rgb, ray_o, ray_d, near, far, coord_, mask_at_box = if_nerf_dutils.sample_ray_h36m(
img, msk, K, R, T, can_bounds, self.nrays, self.split)
acc = if_nerf_dutils.get_acc(coord_, msk)
ret = {
'feature': feature,
'coord': coord,
'out_sh': out_sh,
'rgb': rgb,
'ray_o': ray_o,
'ray_d': ray_d,
'near': near,
'far': far,
'acc': acc,
'mask_at_box': mask_at_box
}
R = cv2.Rodrigues(Rh)[0].astype(np.float32)
i = index // self.num_cams
meta = {'bounds': bounds, 'R': R, 'Th': Th, 'i': i, 'cam_ind': cam_ind}
ret.update(meta)
return ret
def __getitem__(self, index):
img_path = os.path.join(self.data_root, self.ims[index])
img = imageio.imread(img_path).astype(np.float32)[..., :3] / 255.
msk = self.get_mask(index)
cam_ind = self.cam_inds[index]
K = np.array(self.cams['K'][cam_ind])
R = np.array(self.cams['R'][cam_ind])
T = np.array(self.cams['T'][cam_ind])
# reduce the image resolution by ratio
H, W = int(img.shape[0] * cfg.ratio), int(img.shape[1] * cfg.ratio)
img = cv2.resize(img, (W, H), interpolation=cv2.INTER_AREA)
msk = cv2.resize(msk, (W, H), interpolation=cv2.INTER_NEAREST)
if cfg.mask_bkgd:
img[msk == 0] = 0
K[:2] = K[:2] * cfg.ratio
feature, coord, out_sh, can_bounds, bounds, Rh, Th, center, rot, trans = self.prepare_input(
index)
if cfg.sample_smpl:
depth_path = os.path.join(self.data_root, 'depth',
self.ims[index])[:-4] + '.npy'
depth = np.load(depth_path)
rgb, ray_o, ray_d, near, far, coord_, mask_at_box = if_nerf_dutils.sample_smpl_ray(
img, msk, depth, K, R, T, self.nrays, self.split)
elif cfg.sample_grid:
# print('sample_grid')
rgb, ray_o, ray_d, near, far, coord_, mask_at_box = if_nerf_dutils.sample_ray_grid(
img, msk, K, R, T, can_bounds, self.nrays, self.split)
else:
rgb, ray_o, ray_d, near, far, coord_, mask_at_box = if_nerf_dutils.sample_ray_h36m(
img, msk, K, R, T, can_bounds, self.nrays, self.split)
acc = if_nerf_dutils.get_acc(coord_, msk)
ret = {
'feature': feature,
'coord': coord,
'out_sh': out_sh,
'rgb': rgb,
'ray_o': ray_o,
'ray_d': ray_d,
'near': near,
'far': far,
'acc': acc,
'mask_at_box': mask_at_box,
'index': index,
}
R = cv2.Rodrigues(Rh)[0].astype(np.float32)
i = index #// self.num_cams
meta = {
'bounds': bounds,
'R': R,
'Th': Th,
'center': center,
'rot': rot,
'trans': trans,
'i': i,
'cam_ind': cam_ind
}
ret.update(meta)
return ret
def __getitem__(self, index):
# get one image from the ims array
img_path = os.path.join(self.data_root, self.ims[index])
# normalize the values to 0~1
img = imageio.imread(img_path).astype(np.float32) / 255.0
# resize the image to 1024, 1024
img = cv2.resize(img, (1024, 1024))
# get the image mask according to the index
msk = self.get_mask(index)
# get the camera index of this image
cam_ind = self.cam_inds[index]
# get the camera intrinsic parameter K (focal length and optical center)
K = np.array(self.cams["K"][cam_ind])
# get the camera intrinsic parameter D
D = np.array(self.cams["D"][cam_ind])
img = cv2.undistort(img, K, D)
msk = cv2.undistort(msk, K, D)
R = np.array(self.cams["R"][cam_ind])
T = np.array(self.cams["T"][cam_ind]) / 1000.0
# reduce the image resolution by ratio
H, W = int(img.shape[0] * cfg.ratio), int(img.shape[1] * cfg.ratio)
img = cv2.resize(img, (W, H), interpolation=cv2.INTER_AREA)
msk = cv2.resize(msk, (W, H), interpolation=cv2.INTER_NEAREST)
if cfg.mask_bkgd:
img[msk == 0] = 0
if cfg.white_bkgd:
img[msk == 0] = 1
K[:2] = K[:2] * cfg.ratio
if self.human in ["CoreView_313", "CoreView_315"]:
i = int(os.path.basename(img_path).split("_")[4])
frame_index = i - 1
else:
i = int(os.path.basename(img_path)[:-4])
frame_index = i
(
feature, # (6890, 6)
coord, #
out_sh,
can_bounds,
bounds,
Rh,
Th,
center,
rot,
trans,
) = self.prepare_input(i)
(
rgb,
ray_o,
ray_d,
near,
far,
coord_,
mask_at_box,
) = if_nerf_dutils.sample_ray_h36m(img, msk, K, R, T, can_bounds,
self.nrays, self.split)
acc = if_nerf_dutils.get_acc(coord_, msk)
ret = {
"feature": feature,
"coord": coord,
"out_sh": out_sh,
"rgb": rgb,
"ray_o": ray_o,
"ray_d": ray_d,
"near": near,
"far": far,
"acc": acc,
"mask_at_box": mask_at_box,
}
R = cv2.Rodrigues(Rh)[0].astype(np.float32)
i = index // self.num_cams
if cfg.test_novel_pose:
i = 0
meta = {
"bounds": bounds,
"R": R,
"Th": Th,
"center": center,
"rot": rot,
"trans": trans,
"i": i,
"frame_index": frame_index,
"cam_ind": cam_ind,
}
ret.update(meta)
return ret