dataset_script.py

import os
import pandas as pd
from pathlib import Path
import datasets

_CITATION = """
@inproceedings{your_neurips_submission,
  title={Multimodal Street-level Place Recognition Dataset},
  author={Ou, Yiwei},
  year={2025},
  booktitle={NeurIPS Datasets and Benchmarks Track}
}
"""

_DESCRIPTION = """
Multimodal Street-level Place Recognition Dataset (Resized version).
This version loads images, videos, and associated annotations for place recognition tasks,
including GPS, camera metadata, and temporal information.
"""

_HOMEPAGE = "https://huggingface.co/datasets/Yiwei-Ou/Multimodal_Street-level_Place_Recognition_Dataset"
_LICENSE = "cc-by-4.0"

class MultimodalPlaceRecognition(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features({
                "image_path": datasets.Value("string"),
                "video_path": datasets.Value("string"),
                "location_code": datasets.Value("string"),
                "spatial_type": datasets.Value("string"),
                "index": datasets.Value("int32"),
                "shop_names": datasets.Value("string"),
                "sign_text": datasets.Value("string"),
                "image_metadata": datasets.Value("string"),
                "video_metadata": datasets.Value("string"),
            }),
            supervised_keys=None,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        archive_path = dl_manager.download_and_extract(
            "https://huggingface.co/datasets/Yiwei-Ou/Multimodal_Street-level_Place_Recognition_Dataset/resolve/main/Annotated_Resized.tar.gz"
        )
        base_dir = os.path.join(archive_path, "03 Annotated_Resized", "Dataset_Full")
        image_dir = os.path.join(base_dir, "Images")
        video_dir = os.path.join(base_dir, "Videos")
        text_dir = os.path.join(base_dir, "Texts")

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "image_dir": image_dir,
                    "video_dir": video_dir,
                    "annotations_path": os.path.join(text_dir, "Annotations.xlsx"),
                    "image_meta_path": os.path.join(text_dir, "Media_Metadata-Images.xlsx"),
                    "video_meta_path": os.path.join(text_dir, "Media_Metadata-Videos.xlsx"),
                },
            )
        ]

    def _generate_examples(self, image_dir, video_dir, annotations_path, image_meta_path, video_meta_path):
        id_ = 0

        annotations_df = pd.read_excel(annotations_path, engine="openpyxl")
        annotations_dict = {
            str(row["Code"]).strip(): {
                "spatial_type": str(row["Type"]).strip(),
                "index": int(row["Index"]),
                "shop_names": str(row["List of Store Names and Signs"]) if not pd.isna(row["List of Store Names and Signs"]) else "",
                "sign_text": "",  # Not explicitly available
            }
            for _, row in annotations_df.iterrows()
        }

        image_meta_df = pd.read_excel(image_meta_path, engine="openpyxl")
        image_meta_dict = {
            str(row["Filename"]).strip(): row.drop("Filename").dropna().to_dict()
            for _, row in image_meta_df.iterrows()
        }

        video_meta_df = pd.read_excel(video_meta_path, engine="openpyxl")
        video_meta_dict = {
            str(row["Filename"]).strip(): row.drop("Filename").dropna().to_dict()
            for _, row in video_meta_df.iterrows()
        }

        for spatial_type in os.listdir(image_dir):
            spatial_path = os.path.join(image_dir, spatial_type)
            if not os.path.isdir(spatial_path):
                continue
            for location_code in os.listdir(spatial_path):
                loc_img_path = os.path.join(spatial_path, location_code)
                if not os.path.isdir(loc_img_path):
                    continue

                loc_vid_path = os.path.join(video_dir, spatial_type, location_code) if os.path.exists(os.path.join(video_dir, spatial_type, location_code)) else None
                vid_files = set(os.listdir(loc_vid_path)) if loc_vid_path else set()

                for file_name in os.listdir(loc_img_path):
                    if file_name.lower().endswith((".jpg", ".jpeg", ".png")):
                        base_name = os.path.splitext(file_name)[0]
                        video_match = [v for v in vid_files if v.startswith(base_name) and v.endswith(".mp4")]
                        video_file = video_match[0] if video_match else ""
                        video_path = os.path.join(loc_vid_path, video_file) if video_file else ""

                        meta = annotations_dict.get(location_code, {
                            "spatial_type": spatial_type,
                            "index": -1,
                            "shop_names": "",
                            "sign_text": "",
                        })

                        img_meta = image_meta_dict.get(file_name, {})
                        vid_meta = video_meta_dict.get(video_file, {}) if video_file else {}

                        yield id_, {
                            "image_path": os.path.join(loc_img_path, file_name),
                            "video_path": video_path,
                            "location_code": location_code,
                            "spatial_type": meta["spatial_type"],
                            "index": meta["index"],
                            "shop_names": meta["shop_names"],
                            "sign_text": meta["sign_text"],
                            "image_metadata": str(img_meta),
                            "video_metadata": str(vid_meta),
                        }
                        id_ += 1