def bitmask2coco_ins_seg(mask_base: str,
config: Config,
nproc: int = NPROC) -> GtType:
"""Converting BDD100K Instance Segmentation Set to COCO format."""
files = list_files(mask_base, suffix=".png")
images: List[ImgType] = []
logger.info("Collecting bitmasks...")
image_id = 0
for file_ in tqdm(files):
image_id += 1
image = ImgType(
id=image_id,
file_name=file_.replace(".png", ".jpg"),
)
images.append(image)
annotations = bitmask2coco_wo_ids_parallel(mask_base, images, nproc)
return GtType(
type="instances",
categories=get_coco_categories(config),
images=images,
annotations=annotations,
)
def bitmask2coco_seg_track(mask_base: str,
config: Config,
nproc: int = NPROC) -> GtType:
"""Converting BDD100K Instance Segmentation Set to COCO format."""
videos: List[VidType] = []
images: List[ImgType] = []
all_files = list_files(mask_base, suffix=".png")
files_list = group_and_sort_files(all_files)
logger.info("Collecting bitmasks...")
video_id, image_id = 0, 0
for files in files_list:
video_name = os.path.split(files[0])[0]
video_id += 1
video = VidType(id=video_id, name=video_name)
videos.append(video)
for frame_id, file_ in tqdm(enumerate(files)):
image_id += 1
image = ImgType(
video_id=video_id,
frame_id=frame_id,
id=image_id,
file_name=file_.replace(".png", ".jpg"),
)
images.append(image)
annotations = bitmask2coco_wo_ids_parallel(mask_base, images, nproc)
return GtType(
type="instances",
categories=get_coco_categories(config),
videos=videos,
images=images,
annotations=annotations,
)
def evaluate_ins_seg(
gt_paths: List[str],
pred_paths: List[str],
pred_score_file: str,
config: Config,
nproc: int = NPROC,
with_logs: bool = True,
) -> DetResult:
"""Load the ground truth and prediction results.
Args:
gt_paths: paths to the ground truth bitmasks.
pred_paths: paths to the prediciton bitmasks.
pred_score_file: path tothe prediction scores.
config: Config instance.
nproc: number of processes.
with_logs: whether to print logs
Returns:
dict: detection metric scores
"""
categories = get_coco_categories(config)
cat_ids = [category["id"] for category in categories]
cat_names = [category["name"] for category in categories]
pred_paths = reorder_preds(gt_paths, pred_paths)
bdd_eval = BDD100KInsSegEval(gt_paths, pred_paths, pred_score_file,
cat_names, nproc)
bdd_eval.params.catIds = cat_ids
if with_logs:
logger.info("evaluating...")
bdd_eval.evaluate()
if with_logs:
logger.info("accumulating...")
bdd_eval.accumulate()
result = bdd_eval.summarize() # pylint: disable=redefined-outer-name
return result
def bdd100k2coco_seg_track(mask_base: str,
frames: List[Frame],
config: Config,
nproc: int = NPROC) -> GtType:
"""Converting BDD100K Segmentation Tracking Set to COCO format."""
video_id, image_id, ann_id = 0, 0, 0
img_shape = config.imageSize
frames_list = group_and_sort(frames)
videos: List[VidType] = []
images: List[ImgType] = []
mask_names: List[str] = []
category_ids_list: List[List[int]] = []
instance_ids_list: List[List[int]] = []
annotations_list: List[List[AnnType]] = []
categories = get_leaf_categories(config.categories)
cat_name2id = {cat.name: i + 1 for i, cat in enumerate(categories)}
logger.info("Collecting annotations...")
for video_anns in tqdm(frames_list):
global_instance_id: int = 1
instance_id_maps: Dict[str, int] = {}
video_name = video_anns[0].videoName
video_id += 1
video = VidType(id=video_id, name=video_name)
videos.append(video)
for image_anns in video_anns:
image_id += 1
if img_shape is None:
if image_anns.size is not None:
img_shape = image_anns.size
else:
raise ValueError("Image shape not defined!")
image = ImgType(
video_id=video_id,
frame_id=image_anns.frameIndex,
id=image_id,
file_name=os.path.join(video_name, image_anns.name),
height=img_shape.height,
width=img_shape.width,
)
if image_anns.url is not None:
image["coco_url"] = image_anns.url
images.append(image)
mask_name = os.path.join(
mask_base,
video_name,
# Bitmask in .png format, image in .jpg format
image_anns.name.replace(".jpg", ".png"),
)
mask_names.append(mask_name)
category_ids: List[int] = []
instance_ids: List[int] = []
annotations: List[AnnType] = []
for label in image_anns.labels:
if label.poly2d is None:
continue
if label.category not in cat_name2id:
continue
ann_id += 1
instance_id, global_instance_id = get_bdd100k_instance_id(
instance_id_maps, global_instance_id, label.id)
category_id = cat_name2id[label.category]
annotation = AnnType(
id=ann_id,
image_id=image_id,
instance_id=instance_id,
category_id=category_id,
scalabel_id=label.id,
iscrowd=int(check_crowd(label) or check_ignored(label)),
ignore=0,
)
category_ids.append(category_id)
instance_ids.append(instance_id)
annotations.append(annotation)
category_ids_list.append(category_ids)
instance_ids_list.append(instance_ids)
annotations_list.append(annotations)
annotations = bitmask2coco_with_ids_parallel(
annotations_list,
mask_names,
category_ids_list,
instance_ids_list,
nproc,
)
return GtType(
type="instances",
categories=get_coco_categories(config),
videos=videos,
images=images,
annotations=annotations,
)