pascal_voc.py

import os
import logging
import datasets
import xml.etree.ElementTree as ET
from PIL import Image
from collections import defaultdict

_CITATION = """
PASCAL_VOC
"""

_DESCRIPTION = """
PASCAL_VOC
"""

_URLS = {
    "voc2007": "voc2007.tar.gz",
    "voc2012": "voc2012.tar.gz",
}

# fmt: off
CLASS_INFOS = [
    #  class name                id      train          color
    (  'aeroplane'             ,  0 ,        0 , ( 128,   0,   0) ),
    (  'bicycle'               ,  1 ,        1 , (   0, 128,   0) ),
    (  'bird'                  ,  2 ,        2 , ( 128, 128,   0) ),
    (  'boat'                  ,  3 ,        3 , (   0,   0, 128) ),
    (  'bottle'                ,  4 ,        4 , ( 128,   0, 128) ),
    (  'bus'                   ,  5 ,        5 , (   0, 128, 128) ),
    (  'car'                   ,  6 ,        6 , ( 128, 128, 128) ),
    (  'cat'                   ,  7 ,        7 , (  64,   0,   0) ),
    (  'chair'                 ,  8 ,        8 , ( 192,   0,   0) ),
    (  'cow'                   ,  9 ,        9 , (  64, 128,   0) ),
    (  'diningtable'           , 10 ,       10 , ( 192, 128,   0) ),
    (  'dog'                   , 11 ,       11 , (  64,   0, 128) ),
    (  'horse'                 , 12 ,       12 , ( 192,   0, 128) ),
    (  'motorbike'             , 13 ,       13 , (  64, 128, 128) ),
    (  'person'                , 14 ,       14 , ( 192, 128, 128) ),
    (  'pottedplant'           , 15 ,       15 , (   0,  64,   0) ),
    (  'sheep'                 , 16 ,       16 , ( 128,  64,   0) ),
    (  'sofa'                  , 17 ,       17 , (   0, 192,   0) ),
    (  'train'                 , 18 ,       18 , ( 128, 192,   0) ),
    (  'tvmonitor'             , 19 ,       19 , (   0,  64, 128) ),
    (  'background'            , 20 ,       20 , (   0,   0,   0) ),
    (  'borderingregion'       , 255,       21 , ( 224, 224, 192) ),
]

ACTION_INFOS = [
    #  class name                id
    (  'phoning'               ,  0 ),
    (  'playinginstrument'     ,  1 ),
    (  'reading'               ,  2 ),
    (  'ridingbike'            ,  3 ),
    (  'ridinghorse'           ,  4 ),
    (  'running'               ,  5 ),
    (  'takingphoto'           ,  6 ),
    (  'usingcomputer'         ,  7 ),
    (  'walking'               ,  8 ),
    (  'jumping'               ,  9 ),
    (  'other'                 , 10 ),
]

LAYOUT_INFOS = [
    #  class name                id
    (  'Frontal'               ,  0 ),
    (  'Left'                  ,  1 ),
    (  'Rear'                  ,  2 ),
    (  'Right'                 ,  3 ),
    (  'Unspecified'           ,  4 ),
]

# fmt: on

CLASS_NAMES = [CLASS_INFO[0] for CLASS_INFO in CLASS_INFOS]
ACTION_NAMES = [ACTION_INFO[0] for ACTION_INFO in ACTION_INFOS]
LAYOUT_NAMES = [LAYOUT_INFO[0] for LAYOUT_INFO in LAYOUT_INFOS]

CLASS_DICT = {CLASS_INFO[0]: CLASS_INFO[2] for CLASS_INFO in CLASS_INFOS}
ACTION_DICT = {ACTION_INFO[0]: ACTION_INFO[1] for ACTION_INFO in ACTION_INFOS}
LAYOUT_DICT = {LAYOUT_INFO[0]: LAYOUT_INFO[1] for LAYOUT_INFO in LAYOUT_INFOS}

# datasets.Features

action_features = datasets.Features(
    {
        "id": datasets.Value("int32"),
        "image": datasets.features.Image(),
        "height": datasets.Value("int32"),
        "width": datasets.Value("int32"),
        "classes": datasets.features.Sequence(datasets.Value("int32")),
        "objects": datasets.features.Sequence(
            {
                "bboxes": datasets.Sequence(datasets.Value("float32")),
                "classes": datasets.features.ClassLabel(names=ACTION_NAMES),
                "difficult": datasets.Value("int32"),
            }
        ),
    }
)

layout_features = datasets.Features(
    {
        "id": datasets.Value("int32"),
        "image": datasets.features.Image(),
        "height": datasets.Value("int32"),
        "width": datasets.Value("int32"),
        "classes": datasets.features.Sequence(datasets.Value("int32")),
        "objects": datasets.features.Sequence(
            {
                "bboxes": datasets.Sequence(datasets.Value("float32")),
                "classes": datasets.features.ClassLabel(names=LAYOUT_NAMES),
                "difficult": datasets.Value("int32"),
            }
        ),
    }
)

main_features = datasets.Features(
    {
        "id": datasets.Value("int32"),
        "image": datasets.features.Image(),
        "height": datasets.Value("int32"),
        "width": datasets.Value("int32"),
        "classes": datasets.features.Sequence(datasets.Value("int32")),
        "objects": datasets.features.Sequence(
            {
                "bboxes": datasets.Sequence(datasets.Value("float32")),
                "classes": datasets.features.ClassLabel(names=CLASS_NAMES),
                "difficult": datasets.Value("int32"),
            }
        ),
    }
)

segmentation_features = datasets.Features(
    {
        "id": datasets.Value("int32"),
        "image": datasets.features.Image(),
        "height": datasets.Value("int32"),
        "width": datasets.Value("int32"),
        "classes": datasets.features.Sequence(datasets.Value("int32")),
        "class_gt_image": datasets.features.Image(),
        "object_gt_image": datasets.features.Image(),
    }
)

_DATASET_FEATURES = {
    "action": action_features,
    "layout": layout_features,
    "main": main_features,
    "segmentation": segmentation_features,
}


def get_main_classes(data_folder):
    class_infos = defaultdict(set)
    class_folder = os.path.join(data_folder, "ImageSets", "Main")
    for f in os.listdir(class_folder):
        if not f.endswith(".txt") or len(f.split("_")) != 2:
            continue
        lines = open(os.path.join(class_folder, f), "r").read().split("\n")
        name = f.split("_")[0]
        for line in lines:
            spans = line.strip().split(" ")
            spans = list(filter(lambda x: x.strip() != "", spans))
            if len(spans) != 2 or int(spans[1]) != 1:
                continue
            class_infos[spans[0]].add(name)
    return class_infos


def get_annotation(data_folder):
    anno_infos = dict()
    anno_folder = os.path.join(data_folder, "Annotations")
    for f in os.listdir(anno_folder):
        if not f.endswith(".xml"):
            continue
        anno_file = os.path.join(anno_folder, f)
        anno_tree = ET.parse(anno_file)
        objects = []
        for obj in anno_tree.findall("./object"):
            info = {
                "class": obj.findall("./name")[0].text,
                "bbox": [
                    int(float(obj.findall("./bndbox/xmin")[0].text)),
                    int(float(obj.findall("./bndbox/ymin")[0].text)),
                    int(float(obj.findall("./bndbox/xmax")[0].text)),
                    int(float(obj.findall("./bndbox/ymax")[0].text)),
                ],
            }

            if obj.findall("./pose"):
                info["pose"] = obj.findall("./pose")[0].text
            if obj.findall("./truncated"):
                info["truncated"] = int(obj.findall("./truncated")[0].text)
            if obj.findall("./difficult"):
                info["difficult"] = int(obj.findall("./difficult")[0].text)
            else:
                info["difficult"] = 0
            if obj.findall("./occluded"):
                info["occluded"] = int(obj.findall("./occluded")[0].text)

            if obj.findall("./actions"):
                info["action"] = [
                    action.tag
                    for action in obj.findall("./actions/")
                    if int(action.text) == 1
                ][0]

            objects.append(info)
        anno_info = {
            "image": anno_tree.findall("./filename")[0].text,
            "height": int(anno_tree.findall("./size/height")[0].text),
            "width": int(anno_tree.findall("./size/width")[0].text),
            "segmented": int(anno_tree.findall("./segmented")[0].text),
            "objects": objects,
        }
        stem, suffix = os.path.splitext(f)
        anno_infos[stem] = anno_info

    return anno_infos


class PASCALConfig(datasets.BuilderConfig):
    def __init__(self, data_name, task_name, **kwargs):
        """

        Args:
          **kwargs: keyword arguments forwarded to super.
        """
        super().__init__(**kwargs)
        assert data_name in ["voc2007", "voc2012"] and task_name in [
            "action",
            "layout",
            "main",
            "segmentation",
        ]
        assert not (data_name == "voc2007" and task_name == "action")
        self.data_name = data_name
        self.task_name = task_name


class PASCALDataset(datasets.GeneratorBasedBuilder):

    BUILDER_CONFIGS = [
        PASCALConfig(
            name="voc2007_layout",
            version=datasets.Version("1.0.0", ""),
            description="voc2007 layout dataset",
            data_name="voc2007",
            task_name="layout",
        ),
        PASCALConfig(
            name="voc2007_main",
            version=datasets.Version("1.0.0", ""),
            description="voc2007 main dataset",
            data_name="voc2007",
            task_name="main",
        ),
        PASCALConfig(
            name="voc2007_segmentation",
            version=datasets.Version("1.0.0", ""),
            description="voc2007 segmentation dataset",
            data_name="voc2007",
            task_name="segmentation",
        ),
        PASCALConfig(
            name="voc2012_action",
            version=datasets.Version("1.0.0", ""),
            description="voc2012 action dataset",
            data_name="voc2012",
            task_name="action",
        ),
        PASCALConfig(
            name="voc2012_layout",
            version=datasets.Version("1.0.0", ""),
            description="voc2012 layout dataset",
            data_name="voc2012",
            task_name="layout",
        ),
        PASCALConfig(
            name="voc2012_main",
            version=datasets.Version("1.0.0", ""),
            description="voc2012 main dataset",
            data_name="voc2012",
            task_name="main",
        ),
        PASCALConfig(
            name="voc2012_segmentation",
            version=datasets.Version("1.0.0", ""),
            description="voc2012 segmentation dataset",
            data_name="voc2012",
            task_name="segmentation",
        ),
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=_DATASET_FEATURES[self.config.task_name],
            # No default supervised_keys (as we have to pass both question
            # and context as input).
            supervised_keys=None,
            homepage="https://fuliucansheng.github.io/",
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        downloaded_files = dl_manager.download_and_extract(_URLS[self.config.data_name])

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"filepath": downloaded_files, "split": "train"},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"filepath": downloaded_files, "split": "val"},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"filepath": downloaded_files, "split": "test"},
            ),
        ]

    def _generate_examples(self, filepath, split):
        """This function returns the examples in the raw (text) form."""
        logging.info("generating examples from = %s, split = %s", filepath, split)
        data_folder = os.path.join(filepath, os.listdir(filepath)[0])
        anno_infos = get_annotation(data_folder)
        class_infos = get_main_classes(data_folder)
        data_file = os.path.join(
            data_folder, "ImageSets", self.config.task_name.capitalize(), f"{split}.txt"
        )
        with open(data_file, encoding="utf-8") as f:
            for id_, line in enumerate(f):
                line = line.strip()
                if line.count(" ") > 0:
                    line = line.split(" ")[0]
                anno_info = anno_infos.get(line)
                if anno_info is None:
                    continue

                image = os.path.join(data_folder, "JPEGImages", anno_info["image"])
                if not os.path.exists(image):
                    continue

                classes = (
                    [CLASS_DICT[c] for c in class_infos.get(line.strip())]
                    if line.strip() in class_infos
                    else []
                )

                example = {
                    "id": id_,
                    "image": Image.open(os.path.abspath(image)),
                    "height": anno_info["height"],
                    "width": anno_info["width"],
                    "classes": classes,
                }

                objects_info = anno_info["objects"]

                if self.config.task_name == "action":
                    example["objects"] = [
                        {
                            "bboxes": object_info["bbox"],
                            "classes": object_info["action"],
                            "difficult": object_info["difficult"],
                        }
                        for object_info in objects_info
                        if "action" in object_info
                    ]
                    if len(example["objects"]) == 0 and split != "test":
                        continue

                if self.config.task_name == "layout":
                    example["objects"] = [
                        {"bboxes": object_info["bbox"], "classes": object_info["pose"], "difficult": object_info["difficult"],}
                        for object_info in objects_info
                        if "pose" in object_info
                    ]
                    if len(example["objects"]) == 0 and split != "test":
                        continue

                if self.config.task_name == "main":
                    example["objects"] = [
                        {"bboxes": object_info["bbox"], "classes": object_info["class"], "difficult": object_info["difficult"],}
                        for object_info in objects_info
                        if "class" in object_info
                    ]
                    if len(example["objects"]) == 0 and split != "test":
                        continue

                if self.config.task_name == "segmentation":
                    example["class_gt_image"] = Image.open(os.path.abspath(
                        os.path.join(
                            data_folder,
                            "SegmentationClass",
                            anno_info["image"].replace(".jpg", ".png"),
                        )
                    ))
                    example["object_gt_image"] = Image.open(os.path.abspath(
                        os.path.join(
                            data_folder,
                            "SegmentationObject",
                            anno_info["image"].replace(".jpg", ".png"),
                        )
                    ))

                yield id_, example