def __getitem__(self, index):
processed_entry = self.get_processed_entry(self._entry_list[index])
# Do normalization on images
from mankey.utils.imgproc import rgb_image_normalize, depth_image_normalize
# The randomization on rgb
color_aug_scale = self._get_color_randomization_parameter()
normalized_rgb = rgb_image_normalize(processed_entry.cropped_rgb,
self._config.rgb_mean,
color_aug_scale)
#do gray scale here
normalized_gray = np.dot(np.transpose(normalized_rgb, (1, 2, 0)),
[0.2989, 0.5870, 0.1140])
normalized_gray = np.expand_dims(normalized_gray, axis=0) # 1xhxw
gray_channels, height, width = normalized_gray.shape
# Check the total size of tensor
tensor_channels = gray_channels
if processed_entry.has_depth:
tensor_channels += 1
# Construct the tensor
stacked_tensor = np.zeros(shape=(tensor_channels, height, width),
dtype=np.float32)
stacked_tensor[0:gray_channels, :, :] = normalized_gray
# Process other channels
channel_offset = gray_channels
if processed_entry.has_depth:
# The depth should not be randomized
normalized_depth = depth_image_normalize(
processed_entry.cropped_depth, self._config.depth_image_clip,
self._config.depth_image_mean, self._config.depth_image_scale)
stacked_tensor[channel_offset, :, :] = normalized_depth
channel_offset += 1
# Do scale on keypoint xy and depth
normalized_keypoint_xy_depth = processed_entry.keypoint_xy_depth.copy()
normalized_keypoint_xy_depth[0, :] = (
processed_entry.keypoint_xy_depth[0, :] / float(width)) - 0.5
normalized_keypoint_xy_depth[1, :] = (
processed_entry.keypoint_xy_depth[1, :] / float(height)) - 0.5
normalized_keypoint_xy_depth[2, :] = \
(processed_entry.keypoint_xy_depth[2, :] - self._config.depth_image_mean) / float(self._config.depth_image_scale)
normalized_keypoint_xy_depth = np.transpose(
normalized_keypoint_xy_depth, (1, 0))
# OK
validity = np.transpose(processed_entry.keypoint_validity, (1, 0))
return {
parameter.rgbd_image_key:
stacked_tensor,
parameter.keypoint_xyd_key:
normalized_keypoint_xy_depth.astype(np.float32),
parameter.keypoint_validity_key:
validity.astype(np.float32),
parameter.target_heatmap_key:
processed_entry.target_heatmap.astype(np.float32)
}
def proc_input_img_internal(
rgb, depth,
is_path_input, # type: bool
bbox_topleft, # type: PixelCoord
bbox_bottomright, # type: PixelCoord
):
"""
The worker for image processing.
:param rgb: numpy ndarray if is_path_input==False, str if is_path_input==True
:param depth: The same as rgb
:param is_path_input: binary flag for whether the input is path or actual image
:param bbox_topleft: Tight bounding box by maskrcnn
:param bbox_bottomright: Tight bounding box by maskrcnn
:return:
"""
# Get the image and crop them using tight bounding box
if is_path_input:
warped_rgb, bbox2patch = imgproc.get_bbox_cropped_image_path(
rgb, True,
bbox_topleft, bbox_bottomright,
patch_width=parameter.default_patch_size_input, patch_height=parameter.default_patch_size_input,
bbox_scale=parameter.bbox_scale)
warped_depth, _ = imgproc.get_bbox_cropped_image_path(
depth, False,
bbox_topleft, bbox_bottomright,
patch_width=parameter.default_patch_size_input, patch_height=parameter.default_patch_size_input,
bbox_scale=parameter.bbox_scale)
else: # Image input
warped_rgb, bbox2patch = imgproc.get_bbox_cropped_image_raw(
rgb, True,
bbox_topleft, bbox_bottomright,
patch_width=parameter.default_patch_size_input, patch_height=parameter.default_patch_size_input,
bbox_scale=parameter.bbox_scale)
warped_depth, _ = imgproc.get_bbox_cropped_image_raw(
depth, False,
bbox_topleft, bbox_bottomright,
patch_width=parameter.default_patch_size_input, patch_height=parameter.default_patch_size_input,
bbox_scale=parameter.bbox_scale)
# Perform normalization
normalized_rgb = imgproc.rgb_image_normalize(warped_rgb, parameter.rgb_mean, [1.0, 1.0, 1.0])
normalized_depth = imgproc.depth_image_normalize(
warped_depth,
parameter.depth_image_clip,
parameter.depth_image_mean,
parameter.depth_image_scale)
# Construct the tensor
channels, height, width = normalized_rgb.shape
stacked_rgbd = np.zeros(shape=(channels + 1, height, width), dtype=np.float32)
stacked_rgbd[0:3, :, :] = normalized_rgb
stacked_rgbd[3, :, :] = normalized_depth
# OK
imgproc_out = ImageProcOut()
imgproc_out.stacked_rgbd = stacked_rgbd
imgproc_out.bbox2patch = bbox2patch
imgproc_out.warped_rgb = warped_rgb
imgproc_out.warped_depth = warped_depth
return imgproc_out
def __getitem__(self, index):
processed_entry = self.get_processed_entry(self._entry_list[index])
# Do normalization on images
from mankey.utils.imgproc import rgb_image_normalize, depth_image_normalize
# The randomization on rgb
color_aug_scale = self._get_color_randomization_parameter()
normalized_rgb = rgb_image_normalize(processed_entry.cropped_rgb, self._config.rgb_mean, color_aug_scale)
rgb_channels, height, width = normalized_rgb.shape
# Check the total size of tensor
tensor_channels = rgb_channels
if processed_entry.has_depth:
tensor_channels += 1
# Construct the tensor
stacked_tensor = np.zeros(shape=(tensor_channels, height, width), dtype=np.float32)
stacked_tensor[0:rgb_channels, :, :] = normalized_rgb
# Process other channels
channel_offset = rgb_channels
if processed_entry.has_depth:
# The depth should not be randomized
normalized_depth = depth_image_normalize(
processed_entry.cropped_depth,
self._config.depth_image_clip,
self._config.depth_image_mean,
self._config.depth_image_scale)
stacked_tensor[channel_offset, :, :] = normalized_depth
channel_offset += 1
# Do scale on keypoint xy and depth
normalized_keypoint_xy_depth = processed_entry.keypoint_xy_depth.copy()
normalized_keypoint_xy_depth[0, :] = (processed_entry.keypoint_xy_depth[0, :] / float(width)) -0.5
normalized_keypoint_xy_depth[1, :] = (processed_entry.keypoint_xy_depth[1, :] / float(height)) -0.5
normalized_keypoint_xy_depth[2, :] = \
(processed_entry.keypoint_xy_depth[2, :] - self._config.depth_image_mean) / float(self._config.depth_image_scale)
normalized_keypoint_xy_depth = np.transpose(normalized_keypoint_xy_depth, (1, 0))
# OK
validity = np.transpose(processed_entry.keypoint_validity, (1, 0))
return {
parameter.rgbd_image_key: stacked_tensor,
parameter.keypoint_xyd_key: normalized_keypoint_xy_depth.astype(np.float32),
parameter.keypoint_validity_key: validity.astype(np.float32),
parameter.target_heatmap_key: processed_entry.target_heatmap.astype(np.float32),
parameter.delta_rot_key: processed_entry.delta_rotation_matrix.astype(np.float32),
parameter.delta_rot_cls_key: processed_entry.delta_rot_cls.astype(np.int),
parameter.delta_xyz_key: processed_entry.delta_translation.astype(np.float32),
parameter.gripper_pose_key: processed_entry.gripper_pose.astype(np.float32),
parameter.step_size_key: processed_entry.step_size.astype(np.float32)
}
