def evaluate_det(
ann_file: Union[str, List[Frame]],
pred_file: Union[str, List[Frame]],
cfg_path: str,
out_dir: str = "none",
nproc: int = 4,
) -> Dict[str, float]:
"""Load the ground truth and prediction results.
Args:
ann_file: path to the ground truth annotations. "*.json"
pred_file: path to the prediciton results in BDD format. "*.json"
cfg_path: path to the config file
out_dir: output_directory
nproc: processes number for loading jsons
Returns:
dict: detection metric scores
"""
# Convert the annotation file to COCO format
if isinstance(ann_file, str):
ann_frames = load(ann_file, nproc)
else:
ann_frames = ann_file
ann_frames = sorted(ann_frames, key=lambda frame: frame.name)
categories, name_mapping, ignore_mapping = load_coco_config(
mode="det",
filepath=cfg_path,
)
ann_coco = scalabel2coco_detection(SHAPE, ann_frames, categories,
name_mapping, ignore_mapping)
coco_gt = COCOV2(None, ann_coco)
# Load results and convert the predictions
if isinstance(pred_file, str):
pred_frames = load(pred_file, nproc)
else:
pred_frames = pred_file
pred_frames = sorted(pred_frames, key=lambda frame: frame.name)
pred_res = scalabel2coco_detection(SHAPE, pred_frames, categories,
name_mapping,
ignore_mapping)["annotations"]
coco_dt = coco_gt.loadRes(pred_res)
cat_ids = coco_dt.getCatIds()
cat_names = [cat["name"] for cat in coco_dt.loadCats(cat_ids)]
img_ids = sorted(coco_gt.getImgIds())
ann_type = "bbox"
coco_eval = COCOeval(coco_gt, coco_dt, ann_type)
coco_eval.params.imgIds = img_ids
return evaluate_workflow(coco_eval, cat_ids, cat_names, out_dir)
def main() -> None:
"""Main function."""
args = parse_args()
assert args.mode in ["det", "box_track", "ins_seg", "seg_track"]
categories, name_mapping, ignore_mapping = load_coco_config(
mode=args.mode,
filepath=args.config,
ignore_as_class=args.ignore_as_class,
)
shape = (args.height, args.width)
if args.only_mask:
assert args.mode in ["ins_seg", "seg_track"]
convert_function = dict(
ins_seg=bitmask2coco_ins_seg,
seg_track=bitmask2coco_seg_track,
)[args.mode]
coco = convert_function(
args.label,
shape,
list_files(args.label, suffix=".png"),
categories,
args.mask_mode,
args.nproc,
)
else:
if args.mode in ["det", "box_track"]:
convert_func = dict(
det=scalabel2coco_detection,
box_track=scalabel2coco_box_track,
)[args.mode]
else:
convert_func = partial(
dict(
ins_seg=bdd100k2coco_ins_seg,
seg_track=bdd100k2coco_seg_track,
)[args.mode],
mask_base=args.mask_base,
mask_mode=args.mask_mode,
nproc=args.nproc,
)
frames = start_converting(args)
coco = convert_func(
shape=shape,
frames=frames,
categories=categories,
name_mapping=name_mapping,
ignore_mapping=ignore_mapping,
ignore_as_class=args.ignore_as_class,
remove_ignore=args.remove_ignore,
)
logger.info("Saving converted annotations to disk...")
with open(args.output, "w") as f:
json.dump(coco, f)
logger.info("Finished!")
def evaluate_ins_seg(
ann_base: str,
pred_base: str,
pred_score_file: str,
cfg_path: str,
out_dir: str = "none",
) -> Dict[str, float]:
"""Load the ground truth and prediction results.
Args:
ann_base: path to the ground truth bitmasks folder.
pred_base: path to the prediciton bitmasks folder.
pred_score_file: path tothe prediction scores.
cfg_path: path to the config file.
out_dir: output_directory.
Returns:
dict: detection metric scores
"""
categories, _, _ = load_coco_config("ins_seg", cfg_path)
bdd_eval = BDDInsSegEval(ann_base, pred_base, pred_score_file)
cat_ids = [int(category["id"]) for category in categories]
cat_names = [str(category["name"]) for category in categories]
return evaluate_workflow(bdd_eval, cat_ids, cat_names, out_dir)
def segtrack_to_bitmasks(
frames: List[Frame],
out_base: str,
ignore_as_class: bool = False,
remove_ignore: bool = False,
nproc: int = 4,
) -> None:
"""Converting segmentation tracking poly2d to bitmasks."""
frames_list = group_and_sort(frames)
categories, name_mapping, ignore_mapping = load_coco_config(
mode="track",
filepath=DEFAULT_COCO_CONFIG,
ignore_as_class=ignore_as_class,
)
out_paths: List[str] = []
colors_list: List[List[np.ndarray]] = []
poly2ds_list: List[List[List[Poly2D]]] = []
logger.info("Preparing annotations for SegTrack to Bitmasks")
for video_anns in tqdm(frames_list):
global_instance_id: int = 1
instance_id_maps: Dict[str, int] = dict()
video_name = video_anns[0].video_name
out_dir = os.path.join(out_base, video_name)
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
for image_anns in video_anns:
# Bitmask in .png format
image_name = image_anns.name.replace(".jpg", ".png")
image_name = os.path.split(image_name)[-1]
out_path = os.path.join(out_dir, image_name)
out_paths.append(out_path)
colors: List[np.ndarray] = []
poly2ds: List[List[Poly2D]] = []
colors_list.append(colors)
poly2ds_list.append(poly2ds)
labels_ = image_anns.labels
if labels_ is None or len(labels_) == 0:
continue
# Scores higher, rendering later
if labels_[0].score is not None:
labels_ = sorted(labels_, key=lambda label: float(label.score))
for label in labels_:
if label.poly2d is None:
continue
category_ignored, category_id = process_category(
label.category,
categories,
name_mapping,
ignore_mapping,
ignore_as_class=ignore_as_class,
)
if category_ignored and remove_ignore:
continue
instance_id, global_instance_id = get_bdd100k_instance_id(
instance_id_maps, global_instance_id, str(label.id))
color = set_instance_color(label, category_id, instance_id,
category_ignored)
colors.append(color)
poly2ds.append(label.poly2d)
logger.info("Start Conversion for SegTrack to Bitmasks")
frames_to_masks(nproc, out_paths, colors_list, poly2ds_list)
def insseg_to_bitmasks(
frames: List[Frame],
out_base: str,
ignore_as_class: bool = False,
remove_ignore: bool = False,
nproc: int = 4,
) -> None:
"""Converting instance segmentation poly2d to bitmasks."""
os.makedirs(out_base, exist_ok=True)
categories, name_mapping, ignore_mapping = load_coco_config(
mode="track",
filepath=DEFAULT_COCO_CONFIG,
ignore_as_class=ignore_as_class,
)
out_paths: List[str] = []
colors_list: List[List[np.ndarray]] = []
poly2ds_list: List[List[List[Poly2D]]] = []
logger.info("Preparing annotations for InsSeg to Bitmasks")
for image_anns in tqdm(frames):
ann_id = 0
# Bitmask in .png format
image_name = image_anns.name.replace(".jpg", ".png")
image_name = os.path.split(image_name)[-1]
out_path = os.path.join(out_base, image_name)
out_paths.append(out_path)
colors: List[np.ndarray] = []
poly2ds: List[List[Poly2D]] = []
colors_list.append(colors)
poly2ds_list.append(poly2ds)
labels_ = image_anns.labels
if labels_ is None or len(labels_) == 0:
continue
# Scores higher, rendering later
if labels_[0].score is not None:
labels_ = sorted(labels_, key=lambda label: float(label.score))
for label in labels_:
if label.poly2d is None:
continue
category_ignored, category_id = process_category(
label.category,
categories,
name_mapping,
ignore_mapping,
ignore_as_class=ignore_as_class,
)
if remove_ignore and category_ignored:
continue
ann_id += 1
color = set_instance_color(label, category_id, ann_id,
category_ignored)
colors.append(color)
poly2ds.append(label.poly2d)
logger.info("Start conversion for InsSeg to Bitmasks")
frames_to_masks(nproc, out_paths, colors_list, poly2ds_list)