EMT.py

"""EMT dataset."""

import os
import datasets

_HOMEPAGE = "https://github.com/AV-Lab/emt-dataset"

_LICENSE = "CC-BY-SA 4.0"

_CITATION = """  
@article{EMTdataset2025,  
      title={EMT: A Visual Multi-Task Benchmark Dataset for Autonomous Driving in the Arab Gulf Region},  
      author={Nadya Abdel Madjid and Murad Mebrahtu and Abdelmoamen Nasser and Bilal Hassan and Naoufel Werghi and Jorge Dias and Majid Khonji},  
      year={2025},  
      eprint={2502.19260},  
      archivePrefix={arXiv},  
      primaryClass={cs.CV},  
      url={https://arxiv.org/abs/2502.19260}  
}  
""" 

_DESCRIPTION = """\
A multi-task dataset for detection, tracking, prediction, and intention prediction. 
This dataset includes 34,386 annotated frames collected over 57 minutes of driving, with annotations for detection and tracking.
"""

# Annotation repository
_ANNOTATION_REPO = "https://huggingface.co/datasets/KuAvLab/EMT/resolve/main/annotations"

# Tar file URLs for images
_TRAIN_IMAGE_ARCHIVE_URL = "https://huggingface.co/datasets/KuAvLab/EMT/resolve/main/train_images.tar.gz"
_TEST_IMAGE_ARCHIVE_URL = "https://huggingface.co/datasets/KuAvLab/EMT/resolve/main/test_images.tar.gz"


class EMT(datasets.GeneratorBasedBuilder):
    """EMT dataset."""

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="full_size",
            description="All images are in their original size.",
            version=datasets.Version("1.0.0"),
        )
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "image": datasets.Image(),
                    "objects": datasets.Sequence(
                        {
                            "bbox": datasets.Sequence(datasets.Value("float32"), length=4),
                            "class_id": datasets.Value("int32"),
                            "track_id": datasets.Value("int32"),
                            "class_name": datasets.Value("string"),
                        }
                    ),
                }
            ),
            supervised_keys=None,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        """Download train/test images and annotations."""

        # Download and extract images
        image_paths = {
            "train": dl_manager.download_and_extract(_TRAIN_IMAGE_ARCHIVE_URL),
            "test": dl_manager.download_and_extract(_TEST_IMAGE_ARCHIVE_URL),
        }
        
        # Download annotations (extracted automatically)
        annotation_paths = {
            "train": dl_manager.download_and_extract(f"{_ANNOTATION_REPO}/train"),
            "test": dl_manager.download_and_extract(f"{_ANNOTATION_REPO}/test"),
        }

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "images": dl_manager.iter_archive(image_paths["train"]),  
                    "annotation_path": annotation_paths["train"],
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "images": dl_manager.iter_archive(image_paths["test"]), 
                    "annotation_path": annotation_paths["test"],
                },
            ),
        ]

    def _generate_examples(self, images, annotation_path):
        """Generate dataset examples by matching images to their corresponding annotations."""
        
        # Load ALL annotations into memory before iterating over images
        annotations = {}

        for ann_file in os.listdir(annotation_path):  # Iterate over all annotation files
            ann_path = os.path.join(annotation_path, ann_file)
            video_name = os.path.splitext(ann_file)[0]  # Extract video folder name

            with open(ann_path, "r", encoding="utf-8") as f:
                for line in f:
                    parts = line.strip().split()
                    if len(parts) < 8:
                        continue

                    frame_id, track_id, class_name = parts[:3]
                    bbox = list(map(float, parts[4:8]))
                    class_id = _GT_OBJECT_CLASSES.get(class_name, -1)
                    img_name = f"{frame_id}.jpg"

                    # Store annotation in a simple dictionary
                    key = f"{video_name}/{img_name}"
                    if key not in annotations:
                        annotations[key] = []
                    
                    annotations[key].append(
                        {
                            "bbox": bbox,
                            "class_id": class_id,
                            "track_id": int(track_id),
                            "class_name": class_name,
                        }
                    )

        # Yield dataset entries
        idx = 0
        for file_path, file_obj in images:
            img_name = os.path.basename(file_path)
            video_name = os.path.basename(os.path.dirname(file_path))
            key = f"{video_name}/{img_name}"  # Match image to preloaded annotations

            if key in annotations:
                yield idx, {
                    "image": {"path": file_path, "bytes": file_obj.read()},
                    "objects": annotations[key],
                }
                idx += 1