def save_prediction_visualization(
pred_folder: str, image_path: str, image_name: str, pred: np.ndarray,
target_img: np.ndarray, id_to_class_name_map: Mapping[int,
str]) -> None:
"""
Args:
- pred_folder
- image_path
- pred
- target_img
- id_to_class_name_map
Returns:
- None
"""
image_name = Path(image_name).stem
mask_save_dir = pred_folder.replace('gray', 'rgb_mask_predictions')
grid_save_fpath = f'{mask_save_dir}/{image_name}.jpg'
rgb_img = cv2_imread_rgb(image_path)
#save_pred_vs_label_7tuple(rgb_img, pred, target_img, self.id_to_class_name_map, grid_save_fpath)
write_six_img_grid_w_embedded_names(rgb_img, pred, target_img,
id_to_class_name_map, grid_save_fpath)
overlaid_save_fpath = f'{mask_save_dir}_overlaid/{image_name}.jpg'
create_leading_fpath_dirs(overlaid_save_fpath)
frame_visualizer = Visualizer(rgb_img, metadata=None)
overlaid_img = frame_visualizer.overlay_instances(
label_map=pred, id_to_class_name_map=id_to_class_name_map)
imageio.imwrite(overlaid_save_fpath, overlaid_img)
def read_resize_write_rgb(old_fpath: str, new_fpath: str, short_side_sz) -> None: """ Args: old_fpath new_fpath """ img_rgb = cv2_imread_rgb(old_fpath) img_rgb_resized = resize_img_by_short_side(img_rgb, short_side_len=short_side_sz, img_type='rgb') cv2.imwrite(new_fpath, img_rgb_resized[:,:,::-1])
def relabel_pair(old_dataroot: str,
new_dataroot: str,
orig_pair: Tuple[str, str],
remapped_pair: Tuple[str, str],
dname: str,
tax_converter: TaxonomyConverter,
segm_to_class: Mapping[int, int],
dataset_colors: Optional[np.ndarray] = None):
"""
No need to copy the RGB files again. We just update the label file paths.
Args:
- old_dataroot:
- new_dataroot:
- orig_pair: Tuple containing relative path to RGB image and label image
- remapped_pair: Tuple containing relative path to RGB image and label image
- label_mapping_arr:
- dataset_colors:
Returns:
- None
"""
_, orig_rel_label_fpath = orig_pair
_, remapped_rel_label_fpath = remapped_pair
old_label_fpath = f'{old_dataroot}/{orig_rel_label_fpath}'
if not os.path.exists(old_label_fpath):
print("Warning: File " + old_label_fpath + " not found!")
return
if dataset_colors is None:
label_img = imageio.imread(old_label_fpath)
else:
# remap from RGB encoded labels to 1-channel class indices
label_img_rgb = cv2_imread_rgb(old_label_fpath)
label_img = rgb_img_to_obj_cls_img(label_img_rgb, dataset_colors)
if not segm_to_class is None:
label_img_id = label_img[:, :, 0] + (label_img[:, :, 1] * 256) + (
label_img[:, :, 2] * 256**2)
label_img = np.ones(label_img.shape[:2],
dtype=np.uint8) * 255 #initialize with unlabeled
for src, dst in segm_to_class.items():
label_img[label_img_id == src] = dst
labels = torch.tensor(label_img, dtype=torch.int64)
remapped_img = tax_converter.transform_label(labels, dname)
new_label_fpath = f'{new_dataroot}/{remapped_rel_label_fpath}'
create_leading_fpath_dirs(new_label_fpath)
remapped_img = remapped_img.numpy().astype(dtype=np.uint8)
imageio.imwrite(new_label_fpath, remapped_img)
def dump_img_masks(highlight_classname: str, split: str, rgb_img_fpath: str,
folder_prefix: str, api: MapillaryMaskDataset):
"""
"""
# print(f'On image {i}/{len(rgb_img_fpaths)}')
rgb_img = cv2_imread_rgb(rgb_img_fpath)
fname_stem = Path(rgb_img_fpath).stem
label_rgb_fpath = f'{api.dataroot}/{split}/labels/{fname_stem}.png'
label_img_rgb = cv2_imread_rgb(label_rgb_fpath)
label_img = api.labelrgb_to_label(label_img_rgb)
present_classnames = [
api.id_to_classname_map[id] for id in np.unique(label_img)
]
if highlight_classname not in present_classnames:
return
instance_img_fpath = f'{api.dataroot}/{split}/instances/{fname_stem}.png'
instance_img = imageio.imread(instance_img_fpath)
instance_ids = np.unique(instance_img)
for instance_id in instance_ids:
instance_mask = (instance_img == instance_id).astype(np.uint8)
is_single_class, sem_class_ids = mask_belongs_single_semantic_class(
instance_mask, label_img)
assert is_single_class
instance_classid = int(sem_class_ids)
instance_classname = api.id_to_classname_map[int(sem_class_ids)]
if instance_classname != highlight_classname: # not in required_class_names:
continue
save_fpath = f'temp_files/{folder_prefix}_{split}_2020_04_18/mapillary_{fname_stem}_{instance_id}.jpg'
save_binary_mask_double(rgb_img,
instance_mask,
save_fpath,
save_to_disk=True)
def read_resize_write_rgb(old_fpath, new_fpath, short_side_sz):
"""
Args:
- old_fpath
- new_fpath
Returns:
- None
"""
img_rgb = cv2_imread_rgb(old_fpath)
img_rgb_resized = resize_img_by_short_side(img_rgb,
short_side_len=short_side_sz,
img_type='rgb')
cv2.imwrite(new_fpath, img_rgb_resized[:, :, ::-1])
def visual_sanitychecks():
"""
Save every 1000th image of each dataset, with classnames embedded.
"""
save_dir = 'temp_files/MSeg_verify'
for dname, d_info in infos.items():
print(f'Writing visual sanity checks for {dname}...')
if dname in [
'coco-panoptic-inst-201', 'mapillary-public66',
'ade20k-151-inst'
]:
continue # is in RGB format and not comparable
id_to_classname_map = get_dataloader_id_to_classname_map(dname)
splits = ['train', 'val']
split_lists = [d_info.trainlist, d_info.vallist]
for split, split_list in zip(splits, split_lists):
pairs = generate_all_img_label_pair_fpaths(d_info.dataroot,
split_list)
# Save 5 examples from each dataset split
step_sz = math.floor(len(pairs) // 5)
for i, (rgb_fpath, label_fpath) in enumerate(pairs[::step_sz]):
print(f'On {i} of {dname}')
rgb_img = cv2_imread_rgb(rgb_fpath)
label_img = imageio.imread(label_fpath)
fname_stem = Path(rgb_fpath).stem
save_fpath = f'{save_dir}/{dname}_{fname_stem}.jpg'
blend_save_fpath = f'{save_dir}/{dname}_{fname_stem}_blended.jpg'
if rgb_img.ndim == 2:
# this image was grayscale
rgb_img = grayscale_to_color(rgb_img)
form_mask_triple_embedded_classnames(rgb_img,
label_img,
id_to_classname_map,
save_fpath,
save_to_disk=True)
frame_visualizer = Visualizer(rgb_img, metadata=None)
output_img = frame_visualizer.overlay_instances(
label_map=label_img,
id_to_class_name_map=id_to_classname_map)
imageio.imwrite(blend_save_fpath, output_img)
def evaluate_predictions(self, save_vis: bool = True) -> None:
""" Calculate accuracy.
Args:
- data_list:
- pred_folder:
Returns:
- None
"""
pred_folder = self.gray_folder
for i, (image_path, target_path) in enumerate(self.data_list):
if self.args.img_name_unique:
image_name = Path(image_path).stem
else:
image_name = get_unique_stem_from_last_k_strs(image_path)
pred = cv2.imread(os.path.join(pred_folder, image_name + '.png'),
cv2.IMREAD_GRAYSCALE)
target_img = imageio.imread(target_path)
target_img = target_img.astype(np.int64)
target_img = self.convert_label_to_pred_taxonomy(target_img)
self.sam.update_metrics_cpu(pred, target_img,
self.num_eval_classes)
if (i + 1) % self.args.vis_freq == 0:
print_str = f'Evaluating {i + 1}/{len(self.data_list)} on image {image_name+".png"},' + \
f' accuracy {self.sam.accuracy:.4f}.'
logger.info(print_str)
if save_vis:
if (i + 1) % self.args.vis_freq == 0:
mask_save_dir = pred_folder.replace(
'gray', 'rgb_mask_predictions')
grid_save_fpath = f'{mask_save_dir}/{image_name}.png'
rgb_img = cv2_imread_rgb(image_path)
save_pred_vs_label_7tuple(rgb_img, pred, target_img,
self.id_to_class_name_map,
grid_save_fpath)
def relabel_pair(old_dataroot: str,
new_dataroot: str,
orig_pair: Tuple[str, str],
remapped_pair: Tuple[str, str],
label_mapping_arr: np.ndarray,
dataset_colors: Optional[np.ndarray] = None):
"""
No need to copy the RGB files again. We just update the label file paths.
Args:
- old_dataroot:
- new_dataroot:
- orig_pair: Tuple containing relative path to RGB image and label image
- remapped_pair: Tuple containing relative path to RGB image and label image
- label_mapping_arr:
- dataset_colors:
Returns:
- None
"""
_, orig_rel_label_fpath = orig_pair
_, remapped_rel_label_fpath = remapped_pair
old_label_fpath = f'{old_dataroot}/{orig_rel_label_fpath}'
if dataset_colors is None:
label_img = imageio.imread(old_label_fpath)
else:
# remap from RGB encoded labels to 1-channel class indices
label_img_rgb = cv2_imread_rgb(old_label_fpath)
label_img = rgb_img_to_obj_cls_img(label_img_rgb, dataset_colors)
remapped_img = map_semantic_img_fast(label_img, label_mapping_arr)
new_label_fpath = f'{new_dataroot}/{remapped_rel_label_fpath}'
create_leading_fpath_dirs(new_label_fpath)
imageio.imwrite(new_label_fpath, remapped_img)
