def find_largest_size(folder_list, downsampling, network_downsampling,
queue_size):
for folder in folder_list:
# Read mask image
undistorted_mask_boundary = cv2.imread(
str(folder / "undistorted_mask.bmp"), cv2.IMREAD_GRAYSCALE)
# Downsample and crop the undistorted mask image
_, start_h, end_h, start_w, end_w = \
utils.downsample_and_crop_mask(undistorted_mask_boundary, downsampling_factor=downsampling,
divide=network_downsampling)
queue_size.put([end_h - start_h, end_w - start_w])
def pre_processing_data(
process_id, folder_list, downsampling, network_downsampling, is_hsv,
inlier_percentage, visible_interval, suggested_h, suggested_w,
queue_clean_point_list, queue_intrinsic_matrix, queue_point_cloud,
queue_mask_boundary, queue_view_indexes_per_point,
queue_selected_indexes, queue_visible_view_indexes, queue_extrinsics,
queue_projection, queue_crop_positions, queue_estimated_scale):
for folder in folder_list:
# We use folder path as the key for dictionaries
# Read undistorted mask image
folder_str = str(folder)
undistorted_mask_boundary = cv2.imread(
str(folder / "undistorted_mask.bmp"), cv2.IMREAD_GRAYSCALE)
# Downsample and crop the undistorted mask image
cropped_downsampled_undistorted_mask_boundary, start_h, end_h, start_w, end_w = \
utils.downsample_and_crop_mask(undistorted_mask_boundary, downsampling_factor=downsampling,
divide=network_downsampling, suggested_h=suggested_h,
suggested_w=suggested_w)
queue_mask_boundary.put(
[folder_str, cropped_downsampled_undistorted_mask_boundary])
queue_crop_positions.put(
[folder_str, [start_h, end_h, start_w, end_w]])
# Read selected image indexes and stride
stride, selected_indexes = utils.read_selected_indexes(folder)
queue_selected_indexes.put([folder_str, selected_indexes])
# Read visible view indexes
visible_view_indexes = utils.read_visible_view_indexes(folder)
queue_visible_view_indexes.put([folder_str, visible_view_indexes])
# Read undistorted camera intrinsics
undistorted_camera_intrinsic_per_view = utils.read_camera_intrinsic_per_view(
folder)
# Downsample and crop the undistorted camera intrinsics
# Assuming for now that camera intrinsics within each clip remains the same
cropped_downsampled_undistorted_intrinsic_matrix = utils.modify_camera_intrinsic_matrix(
undistorted_camera_intrinsic_per_view[0],
start_h=start_h,
start_w=start_w,
downsampling_factor=downsampling)
queue_intrinsic_matrix.put(
[folder_str, cropped_downsampled_undistorted_intrinsic_matrix])
# Read sparse point cloud from SfM
point_cloud = utils.read_point_cloud(str(folder / "structure.ply"))
queue_point_cloud.put([folder_str, point_cloud])
# self.point_cloud_per_seq[folder] = point_cloud
# Read visible view indexes per point
view_indexes_per_point = utils.read_view_indexes_per_point(
folder,
visible_view_indexes=visible_view_indexes,
point_cloud_count=len(point_cloud))
# Update view_indexes_per_point_per_seq with neighborhood frames to increase stability and
# avoid as much occlusion problem as possible
view_indexes_per_point = utils.overlapping_visible_view_indexes_per_point(
view_indexes_per_point, visible_interval)
queue_view_indexes_per_point.put([folder_str, view_indexes_per_point])
# Read pose data for all visible views
poses = utils.read_pose_data(folder)
# Calculate extrinsic and projection matrices
visible_extrinsic_matrices, visible_cropped_downsampled_undistorted_projection_matrices = \
utils.get_extrinsic_matrix_and_projection_matrix(poses,
intrinsic_matrix=
cropped_downsampled_undistorted_intrinsic_matrix,
visible_view_count=len(visible_view_indexes))
queue_extrinsics.put([folder_str, visible_extrinsic_matrices])
queue_projection.put([
folder_str,
visible_cropped_downsampled_undistorted_projection_matrices
])
# Get approximate data global scale to reduce training data imbalance
global_scale = utils.global_scale_estimation(
visible_extrinsic_matrices, point_cloud)
queue_estimated_scale.put([folder_str, global_scale])
visible_cropped_downsampled_imgs = utils.get_color_imgs(
folder,
visible_view_indexes=visible_view_indexes,
start_h=start_h,
start_w=start_w,
end_h=end_h,
end_w=end_w,
downsampling_factor=downsampling,
is_hsv=is_hsv)
# Calculate contaminated point list
clean_point_indicator_array = utils.get_clean_point_list(
imgs=visible_cropped_downsampled_imgs,
point_cloud=point_cloud,
mask_boundary=cropped_downsampled_undistorted_mask_boundary,
inlier_percentage=inlier_percentage,
projection_matrices=
visible_cropped_downsampled_undistorted_projection_matrices,
extrinsic_matrices=visible_extrinsic_matrices,
is_hsv=is_hsv,
view_indexes_per_point=view_indexes_per_point)
queue_clean_point_list.put([folder_str, clean_point_indicator_array])
print("sequence {} finished".format(folder_str))
print("{}th process finished".format(process_id))