import json
from pathlib import Path
import pandas as pd
from tqdm.auto import tqdm
from ...utils import BBOX_COLUMN_NAMES
from ..dataset import Dataset
from .common import get_image_info, parse_annotation_name
BOX_XMIN, BOX_WIDTH, BOX_YMIN, BOX_HEIGHT = BBOX_COLUMN_NAMES
[docs]
def from_crowd_human(
annotation_odgt: Path | str,
images_root: Path | str | None = None,
visible_box: bool = True,
dataset_name: str | None = None,
split: str | None = None,
) -> Dataset:
"""Read a dataset in the format described for CrowdHuman
See https://www.crowdhuman.org/download.html
Args:
annotation_odgt: annotation file using the format described in link given above.
images_root: Folder where images are stored. Note that since annotations IDs are
file names, all images must be at the root of this folder.
If set to None, will use the image folder in
the annotation file's parent folder. Defaults to None.
visible_box: If set to True, will only take the visible bounding box. Otherwise,
will take the whole bounding box, with possibly occluded parts.
dataset_name: If specified, will be the dataset name, used when showing the
dataset or exporting in other formats such as fiftyone. If not specified,
the dataset name will be deduced from the name of the json file.
split: Split value to give to the dataset. If set to None, will try to deduce
it from the annotation file name. Defaults to None.
Raises:
FileNotFoundError: Will raise an error if the images_root folder does not exist
Unlike COCO or Caipy, image size information is not stored in annotations
and thus need to be computed by loading picture's headers.
Returns:
Dataset
"""
annotation_odgt = Path(annotation_odgt)
if images_root is None:
images_root = annotation_odgt.parent / "images"
else:
images_root = Path(images_root)
if not images_root.is_dir():
raise FileNotFoundError(
"Image folder needs to exist as CrowdHuman's annotations don't have image"
" size information"
)
parsed_dataset_name, parsed_split = parse_annotation_name(
annotations_file_path=annotation_odgt
)
if dataset_name is None:
dataset_name = parsed_dataset_name
if split is None:
split = parsed_split
with open(annotation_odgt) as f:
lines = f.read().strip().split("\n")
annots = [json.loads(line) for line in lines]
image_dicts = []
annotation_dicts = []
label_map = {0: "person", 1: "head"}
ids_map = {v: k for k, v in label_map.items()}
for image_id, image_data in enumerate(tqdm(annots)):
relative_path = Path(f"{image_data['ID']}.jpg")
image_path = images_root / relative_path
image_info = get_image_info(image_id, relative_path, image_path)
image_dicts.append(image_info)
current_annotations = []
if "gtboxes" not in image_data:
continue
for annot in image_data["gtboxes"]:
if annot["tag"] == "mask":
continue
if "ignore" in annot["extra"] and annot["extra"]["ignore"] == 1:
continue
*_, box_w, box_h = annot["fbox"]
area = box_w * box_h
*_, vbox_w, vbox_h = annot["vbox"]
visible_share = vbox_w * vbox_h / area
person_annotation = {
**annot["extra"],
"image_id": image_id,
"category_id": ids_map["person"],
"bbox": annot["vbox"] if visible_box else annot["fbox"],
"visible_share": visible_share,
}
current_annotations.append(person_annotation)
if "ignore" in annot["head_attr"] and annot["head_attr"]["ignore"] == 1:
continue
# We don't know how visible each head is
# Sometimes it's occluded, sometimes it's even out of the screen
head_annotation = {
**annot["extra"],
**annot["head_attr"],
"image_id": image_id,
"category_id": ids_map["head"],
"bbox": annot["hbox"],
"visible_share": float("nan"),
}
current_annotations.append(head_annotation)
annotation_dicts.extend(current_annotations)
images = pd.DataFrame.from_records(image_dicts).set_index("id")
images["split"] = split
annotations = pd.DataFrame.from_records(annotation_dicts)
annotations.index.name = "id"
bboxes = pd.DataFrame(list(annotations["bbox"]), index=annotations.index)
for i, name in enumerate(BBOX_COLUMN_NAMES):
annotations[name] = bboxes[i]
annotations["split"] = split
annotations = annotations.drop(["bbox", "ignore"], axis=1, errors="ignore")
annotations["unsure"] = annotations["unsure"].fillna(0)
return Dataset(
images_root=images_root,
images=images,
annotations=annotations,
label_map=label_map,
dataset_name=dataset_name,
).cap_bounding_box_coordinates()
