Urban-ImageNet_Croissant Metadata.json

{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "sc": "https://schema.org/",
    "cr": "http://mlcommons.org/croissant/",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "dct": "http://purl.org/dc/terms/",
    "prov": "http://www.w3.org/ns/prov#",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "containedIn": "cr:containedIn",
    "data": {
      "@id": "cr:data",
      "@type": "@json"
    },
    "dataType": {
      "@id": "cr:dataType",
      "@type": "@vocab"
    },
    "examples": {
      "@id": "cr:examples",
      "@type": "@json"
    },
    "extract": "cr:extract",
    "field": "cr:field",
    "fileObject": "cr:fileObject",
    "fileProperty": "cr:fileProperty",
    "fileSet": "cr:fileSet",
    "format": "cr:format",
    "includes": "cr:includes",
    "isLiveDataset": "cr:isLiveDataset",
    "jsonPath": "cr:jsonPath",
    "key": "cr:key",
    "md5": "cr:md5",
    "recordSet": "cr:recordSet",
    "references": "cr:references",
    "regex": "cr:regex",
    "repeated": "cr:repeated",
    "source": "cr:source",
    "subField": "cr:subField",
    "transform": "cr:transform",
    "value": "cr:value"
  },
  "@type": "sc:Dataset",
  "conformsTo": "http://mlcommons.org/croissant/1.1",
  "name": "Urban-ImageNet",
  "alternateName": "Urban-ImageNet: A Large-Scale Multi-Modal Dataset for Urban Space Perception Benchmarking",
  "url": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet",
  "sameAs": [
    "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet"
  ],
  "license": "https://creativecommons.org/licenses/by-nc/4.0/",
  "usageInfo": "For non-commercial academic research only. Prohibited uses include re-identification, account reconstruction, surveillance, facial recognition, social scoring, law-enforcement targeting, and commercial profiling. Users must respect platform policies and the dataset data-use agreement.",
  "version": "1.0",
  "description": "Urban-ImageNet is a large-scale multimodal dataset and benchmark for urban commercial space perception. It contains over 2 million public Weibo image-text pairs collected from 61 commercial sites in 24 Chinese cities across 2019-2025. The dataset is organized by the HUSIC 10-class taxonomy and supports urban scene classification, category-level and post-level image-text retrieval, instance segmentation, and scaling studies through 1K, 10K, 100K, and full 2M variants.",
  "dateCreated": "2019-01-01",
  "datePublished": "2026-05-05",
  "dateModified": "2026-05-05",
  "inLanguage": [
    "zh",
    "en"
  ],
  "keywords": [
    "urban perception",
    "image classification",
    "image-text retrieval",
    "instance segmentation",
    "social media",
    "Weibo",
    "Chinese cities",
    "urban commercial spaces",
    "multimodal learning",
    "computational urban studies"
  ],
  "creator": [
    {
      "@type": "sc:Organization",
      "name": "Urban-ImageNet Research Team"
    }
  ],
  "publisher": {
    "@type": "sc:Organization",
    "name": "Urban-ImageNet Research Team"
  },
  "spatialCoverage": {
    "@type": "sc:Place",
    "name": "24 Chinese cities and 61 urban commercial sites",
    "geo": {
      "@type": "sc:GeoShape",
      "box": "18.0 73.0 54.0 135.0"
    }
  },
  "temporalCoverage": "2019-01-01/2025-12-31",
  "isAccessibleForFree": true,
  "isLiveDataset": false,
  "distribution": [
    {
      "@type": "cr:FileObject",
      "@id": "1k-sample-zip",
      "name": "1K Dataset-Sample.zip",
      "description": "Compressed 1K sample package for quick inspection and validator-friendly access.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/resolve/main/1K%20Dataset-Sample.zip",
      "encodingFormat": "application/zip",
      "contentSize": "71744171 B",
      "sha256": "ed0dd2c1087ad56c7d5098a3990cb6b871dff0418fd542a4cbe07bff39c6ee2a"
    },
    {
      "@type": "cr:FileSet",
      "@id": "1k-t1-image-files",
      "name": "1K Dataset Task 1 image files",
      "description": "ImageFolder-style JPG files under train/val/test and HUSIC class folders.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/1K%20Dataset/01%20Images%20with%20labels",
      "includes": [
        "train/*/*.jpg",
        "val/*/*.jpg",
        "test/*/*.jpg"
      ],
      "encodingFormat": "image/jpeg"
    },
    {
      "@type": "cr:FileSet",
      "@id": "1k-t2-spreadsheets",
      "name": "1K Dataset Task 2 text-image spreadsheets",
      "description": "Excel files train.xlsx, val.xlsx, and test.xlsx with image-text pair metadata.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/1K%20Dataset/02%20Text-Image%20Pairs",
      "includes": [
        "train.xlsx",
        "val.xlsx",
        "test.xlsx"
      ],
      "encodingFormat": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"
    },
    {
      "@type": "cr:FileSet",
      "@id": "1k-t3-json",
      "name": "1K Dataset Task 3 instance segmentation JSON files",
      "description": "COCO-style JSON annotation files train.json, val.json, and test.json.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/1K%20Dataset/03%20Instance%20Segmentation",
      "includes": [
        "train.json",
        "val.json",
        "test.json"
      ],
      "encodingFormat": "application/json"
    },
    {
      "@type": "cr:FileSet",
      "@id": "10k-t1-image-files",
      "name": "10K Dataset Task 1 image files",
      "description": "ImageFolder-style JPG files under train/val/test and HUSIC class folders.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/10K%20Dataset/01%20Images%20with%20labels",
      "includes": [
        "train/*/*.jpg",
        "val/*/*.jpg",
        "test/*/*.jpg"
      ],
      "encodingFormat": "image/jpeg"
    },
    {
      "@type": "cr:FileSet",
      "@id": "10k-t2-spreadsheets",
      "name": "10K Dataset Task 2 text-image spreadsheets",
      "description": "Excel files train.xlsx, val.xlsx, and test.xlsx with image-text pair metadata.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/10K%20Dataset/02%20Text-Image%20Pairs",
      "includes": [
        "train.xlsx",
        "val.xlsx",
        "test.xlsx"
      ],
      "encodingFormat": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"
    },
    {
      "@type": "cr:FileSet",
      "@id": "10k-t3-json",
      "name": "10K Dataset Task 3 instance segmentation JSON files",
      "description": "COCO-style JSON annotation files train.json, val.json, and test.json.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/10K%20Dataset/03%20Instance%20Segmentation",
      "includes": [
        "train.json",
        "val.json",
        "test.json"
      ],
      "encodingFormat": "application/json"
    },
    {
      "@type": "cr:FileSet",
      "@id": "100k-t1-image-files",
      "name": "100K Dataset Task 1 image files",
      "description": "ImageFolder-style JPG files under train/val/test and HUSIC class folders.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/100K%20Dataset/01%20Images%20with%20labels",
      "includes": [
        "train/*/*.jpg",
        "val/*/*.jpg",
        "test/*/*.jpg"
      ],
      "encodingFormat": "image/jpeg"
    },
    {
      "@type": "cr:FileSet",
      "@id": "100k-t2-spreadsheets",
      "name": "100K Dataset Task 2 text-image spreadsheets",
      "description": "Excel files train.xlsx, val.xlsx, and test.xlsx with image-text pair metadata.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/100K%20Dataset/02%20Text-Image%20Pairs",
      "includes": [
        "train.xlsx",
        "val.xlsx",
        "test.xlsx"
      ],
      "encodingFormat": "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"
    },
    {
      "@type": "cr:FileSet",
      "@id": "100k-t3-json",
      "name": "100K Dataset Task 3 instance segmentation JSON files",
      "description": "COCO-style JSON annotation files train.json, val.json, and test.json.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/100K%20Dataset/03%20Instance%20Segmentation",
      "includes": [
        "train.json",
        "val.json",
        "test.json"
      ],
      "encodingFormat": "application/json"
    },
    {
      "@type": "cr:FileSet",
      "@id": "full-2m-image-files",
      "name": "Full Dataset-2M image files",
      "description": "Flat directory containing all privacy-protected 512 px images in the full unbalanced corpus.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/Full%20Dataset-2M/Images",
      "includes": [
        "*.jpg",
        "*.jpeg"
      ],
      "encodingFormat": "image/jpeg"
    },
    {
      "@type": "cr:FileSet",
      "@id": "full-2m-label-files",
      "name": "Full Dataset-2M label files",
      "description": "CSV label files for semantic classification, text-image pairs, and instance segmentation metadata.",
      "contentUrl": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet/tree/main/Full%20Dataset-2M/Labels",
      "includes": [
        "01 Semantic classification labels.CSV",
        "02 Text-Image Pairs.CSV",
        "03 Instance Segmentation labels.CSV"
      ],
      "encodingFormat": "text/csv"
    }
  ],
  "recordSet": [
    {
      "@type": "cr:RecordSet",
      "@id": "dataset-splits",
      "name": "DatasetSplitRecords",
      "description": "Inline summary of the four Urban-ImageNet scale variants.",
      "key": {
        "@id": "dataset-splits/split_id"
      },
      "field": [
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/split_id",
          "name": "split_id",
          "description": "Stable split identifier.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/name",
          "name": "name",
          "description": "Human-readable split name.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/image_count",
          "name": "image_count",
          "description": "Primary image count.",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/balanced",
          "name": "balanced",
          "description": "Whether the 10 classes are balanced.",
          "dataType": "sc:Boolean"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/predefined_split",
          "name": "predefined_split",
          "description": "Whether train/validation/test partitions are provided.",
          "dataType": "sc:Boolean"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/train_count",
          "name": "train_count",
          "description": "Training image count when predefined.",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/validation_count",
          "name": "validation_count",
          "description": "Validation image count when predefined.",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/test_count",
          "name": "test_count",
          "description": "Test image count when predefined.",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/local_bytes",
          "name": "local_bytes",
          "description": "Local processed package size in bytes when measured.",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "dataset-splits/notes",
          "name": "notes",
          "description": "Usage notes for the split.",
          "dataType": "sc:Text"
        }
      ],
      "data": [
        {
          "dataset-splits/split_id": "1k",
          "dataset-splits/name": "1K Dataset",
          "dataset-splits/image_count": 1000,
          "dataset-splits/balanced": true,
          "dataset-splits/predefined_split": true,
          "dataset-splits/train_count": 800,
          "dataset-splits/validation_count": 100,
          "dataset-splits/test_count": 100,
          "dataset-splits/local_bytes": 91533081,
          "dataset-splits/notes": "Balanced 10-class subset with 100 images per class; includes T1, T2, and T3 annotations."
        },
        {
          "dataset-splits/split_id": "10k",
          "dataset-splits/name": "10K Dataset",
          "dataset-splits/image_count": 10000,
          "dataset-splits/balanced": true,
          "dataset-splits/predefined_split": true,
          "dataset-splits/train_count": 8000,
          "dataset-splits/validation_count": 1000,
          "dataset-splits/test_count": 1000,
          "dataset-splits/local_bytes": 734595582,
          "dataset-splits/notes": "Balanced 10-class subset with 1,000 images per class; includes T1, T2, and T3 annotations."
        },
        {
          "dataset-splits/split_id": "100k",
          "dataset-splits/name": "100K Dataset",
          "dataset-splits/image_count": 100000,
          "dataset-splits/balanced": true,
          "dataset-splits/predefined_split": true,
          "dataset-splits/train_count": 80000,
          "dataset-splits/validation_count": 10000,
          "dataset-splits/test_count": 10000,
          "dataset-splits/local_bytes": 6264915204,
          "dataset-splits/notes": "Balanced benchmark subset with 10,000 images per class; primary split for reported baselines."
        },
        {
          "dataset-splits/split_id": "full-2m",
          "dataset-splits/name": "Full Dataset-2M",
          "dataset-splits/image_count": 2000000,
          "dataset-splits/balanced": false,
          "dataset-splits/predefined_split": false,
          "dataset-splits/train_count": null,
          "dataset-splits/validation_count": null,
          "dataset-splits/test_count": null,
          "dataset-splits/local_bytes": null,
          "dataset-splits/notes": "Full unbalanced corpus with flat Images and Labels folders; users create task-specific splits."
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "husic-classes",
      "name": "HUSICClassRecords",
      "description": "The 10-class HUSIC taxonomy used by all tasks.",
      "key": {
        "@id": "husic-classes/class_id"
      },
      "field": [
        {
          "@type": "cr:Field",
          "@id": "husic-classes/class_id",
          "name": "class_id",
          "description": "Integer HUSIC class identifier.",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "husic-classes/label",
          "name": "label",
          "description": "Full HUSIC label.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "husic-classes/group",
          "name": "group",
          "description": "High-level semantic group.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "husic-classes/definition",
          "name": "definition",
          "description": "Annotation definition.",
          "dataType": "sc:Text"
        }
      ],
      "data": [
        {
          "husic-classes/class_id": 0,
          "husic-classes/label": "Exterior urban spaces with people",
          "husic-classes/group": "Exterior",
          "husic-classes/definition": "Outdoor urban commercial spaces with visible human presence."
        },
        {
          "husic-classes/class_id": 1,
          "husic-classes/label": "Exterior urban spaces without people",
          "husic-classes/group": "Exterior",
          "husic-classes/definition": "Outdoor commercial architecture or public-realm views without visible people."
        },
        {
          "husic-classes/class_id": 2,
          "husic-classes/label": "Interior urban spaces with people",
          "husic-classes/group": "Interior",
          "husic-classes/definition": "Public or semi-public commercial interiors with occupants, shoppers, workers, or event participants."
        },
        {
          "husic-classes/class_id": 3,
          "husic-classes/label": "Interior urban spaces without people",
          "husic-classes/group": "Interior",
          "husic-classes/definition": "Interior commercial spaces focused on spatial design, circulation, fixtures, or displays without people."
        },
        {
          "husic-classes/class_id": 4,
          "husic-classes/label": "Hotel or commercial lodging spaces",
          "husic-classes/group": "Accommodation",
          "husic-classes/definition": "Hotel rooms, serviced apartments, and commercial lodging environments."
        },
        {
          "husic-classes/class_id": 5,
          "husic-classes/label": "Private home interiors",
          "husic-classes/group": "Accommodation",
          "husic-classes/definition": "Private residential interiors connected to the broader commercial-district social-media corpus."
        },
        {
          "husic-classes/class_id": 6,
          "husic-classes/label": "Food or drink items",
          "husic-classes/group": "Consumption",
          "husic-classes/definition": "Food, drinks, dining-table scenes, and restaurant consumption content."
        },
        {
          "husic-classes/class_id": 7,
          "husic-classes/label": "Retail products and merchandise",
          "husic-classes/group": "Consumption",
          "husic-classes/definition": "Product, merchandise, retail shelf, shopping, and display-window content."
        },
        {
          "husic-classes/class_id": 8,
          "husic-classes/label": "Human-centered portrait",
          "husic-classes/group": "Portrait",
          "husic-classes/definition": "Selfies, group photos, portraits, and other images where specific people dominate the composition."
        },
        {
          "husic-classes/class_id": 9,
          "husic-classes/label": "Other non-spatial content",
          "husic-classes/group": "Miscellaneous",
          "husic-classes/definition": "Advertisements, screenshots, memes, maps, QR-code-like content, animals, or other non-spatial material."
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "task-schemas",
      "name": "BenchmarkTaskRecords",
      "description": "Task definitions and ground-truth semantics for T1, T2, and T3.",
      "key": {
        "@id": "task-schemas/task_id"
      },
      "field": [
        {
          "@type": "cr:Field",
          "@id": "task-schemas/task_id",
          "name": "task_id",
          "description": "Task identifier.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "task-schemas/name",
          "name": "name",
          "description": "Task name.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "task-schemas/input",
          "name": "input",
          "description": "Input modality or modalities.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "task-schemas/output",
          "name": "output",
          "description": "Output label, retrieval target, or annotation.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "task-schemas/ground_truth",
          "name": "ground_truth",
          "description": "Ground-truth construction.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "task-schemas/format",
          "name": "format",
          "description": "Released file format.",
          "dataType": "sc:Text"
        }
      ],
      "data": [
        {
          "task-schemas/task_id": "T1",
          "task-schemas/name": "Urban scene semantic classification",
          "task-schemas/input": "Image",
          "task-schemas/output": "One of 10 HUSIC class labels",
          "task-schemas/ground_truth": "Class folder name and Image Label column.",
          "task-schemas/format": "ImageFolder train/val/test directories."
        },
        {
          "task-schemas/task_id": "T2-A",
          "task-schemas/name": "Category-level image-text retrieval",
          "task-schemas/input": "Image or HUSIC label text",
          "task-schemas/output": "Matching HUSIC label text or images in the class",
          "task-schemas/ground_truth": "Image Label column and HUSIC class definitions.",
          "task-schemas/format": "Excel rows joined to images by Image Filename."
        },
        {
          "task-schemas/task_id": "T2-B",
          "task-schemas/name": "Post-level image-text retrieval",
          "task-schemas/input": "Image or original Chinese Post Text",
          "task-schemas/output": "Images attached to the same public post or the corresponding post text",
          "task-schemas/ground_truth": "One-to-many grouping by anonymized post and Image Filename stem.",
          "task-schemas/format": "Excel rows with Post Text and Image Filename join key."
        },
        {
          "task-schemas/task_id": "T3",
          "task-schemas/name": "Instance segmentation",
          "task-schemas/input": "Image",
          "task-schemas/output": "Bounding boxes, COCO RLE masks, detected labels, detection scores",
          "task-schemas/ground_truth": "Grounding DINO + SAM2 pseudo-labels with quality filtering and review.",
          "task-schemas/format": "COCO-style JSON with img_info/images, categories, and annotations."
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "t2-columns",
      "name": "TextImagePairColumnRecords",
      "description": "Column-level schema for train.xlsx, val.xlsx, test.xlsx, and full-corpus CSV text-image pair files.",
      "key": {
        "@id": "t2-columns/column_name"
      },
      "field": [
        {
          "@type": "cr:Field",
          "@id": "t2-columns/column_name",
          "name": "column_name",
          "description": "Excel/CSV column name.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "t2-columns/description",
          "name": "description",
          "description": "Column meaning and use.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "t2-columns/task_role",
          "name": "task_role",
          "description": "Task or metadata role.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "t2-columns/example_value",
          "name": "example_value",
          "description": "English illustrative example; released Post Text remains original Chinese.",
          "dataType": "sc:Text"
        }
      ],
      "data": [
        {
          "t2-columns/column_name": "Image Label",
          "t2-columns/description": "HUSIC class label for the image; used for category-level image-text alignment and T1 labels.",
          "t2-columns/task_role": "T1/T2",
          "t2-columns/example_value": "Exterior urban spaces with people"
        },
        {
          "t2-columns/column_name": "Image Filename",
          "t2-columns/description": "Image stem used to join spreadsheet rows to image files; follows UserID_PostDate_Index.",
          "t2-columns/task_role": "Join key",
          "t2-columns/example_value": "2668383_2020-01-21_0"
        },
        {
          "t2-columns/column_name": "Post ID",
          "t2-columns/description": "Anonymized numerical post identifier.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "4469639217818281"
        },
        {
          "t2-columns/column_name": "User ID",
          "t2-columns/description": "Anonymized numerical user identifier; original usernames are not released.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "2668383"
        },
        {
          "t2-columns/column_name": "Post Time",
          "t2-columns/description": "Original post timestamp in Chinese date-time format.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "2020-01-21 01:00"
        },
        {
          "t2-columns/column_name": "Post Text",
          "t2-columns/description": "Original Chinese Weibo text; used for post-level image-text retrieval.",
          "t2-columns/task_role": "T2",
          "t2-columns/example_value": "Synthetic English illustration: A lively evening walk near a commercial plaza; the released dataset keeps the original Chinese text."
        },
        {
          "t2-columns/column_name": "City",
          "t2-columns/description": "City associated with the location hashtag or commercial site.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "Chengdu"
        },
        {
          "t2-columns/column_name": "Place Tag",
          "t2-columns/description": "Location hashtag or place tag used for collection.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "#Chengdu Taikoo Li"
        },
        {
          "t2-columns/column_name": "Posting Tool",
          "t2-columns/description": "Client or posting-source string after metadata minimization.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "mobile client"
        },
        {
          "t2-columns/column_name": "Mentioned Users",
          "t2-columns/description": "Anonymized or removed mentioned-user field.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": ""
        },
        {
          "t2-columns/column_name": "Extracted Topics",
          "t2-columns/description": "Topic or hashtag terms extracted from the post text.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "city life; shopping district"
        },
        {
          "t2-columns/column_name": "Extracted Locations",
          "t2-columns/description": "Location mentions extracted from the post text.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "Chengdu"
        },
        {
          "t2-columns/column_name": "Like Count",
          "t2-columns/description": "Public engagement count at collection time.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "52"
        },
        {
          "t2-columns/column_name": "Repost Count",
          "t2-columns/description": "Public repost count at collection time.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "13"
        },
        {
          "t2-columns/column_name": "Comment Count",
          "t2-columns/description": "Public comment count at collection time.",
          "t2-columns/task_role": "Metadata",
          "t2-columns/example_value": "19"
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "segmentation-prompts",
      "name": "SegmentationPromptRecords",
      "description": "Per-class prompt vocabularies used to produce Grounding DINO + SAM2 pseudo-labels.",
      "key": {
        "@id": "segmentation-prompts/class_id"
      },
      "field": [
        {
          "@type": "cr:Field",
          "@id": "segmentation-prompts/class_id",
          "name": "class_id",
          "description": "HUSIC class identifier.",
          "dataType": "sc:Integer"
        },
        {
          "@type": "cr:Field",
          "@id": "segmentation-prompts/label",
          "name": "label",
          "description": "HUSIC label.",
          "dataType": "sc:Text"
        },
        {
          "@type": "cr:Field",
          "@id": "segmentation-prompts/prompt",
          "name": "prompt",
          "description": "Per-class Grounding DINO prompt vocabulary.",
          "dataType": "sc:Text"
        }
      ],
      "data": [
        {
          "segmentation-prompts/class_id": 0,
          "segmentation-prompts/label": "Exterior urban spaces with people",
          "segmentation-prompts/prompt": "person . crowd . pedestrian . building facade . lawn . street lamp . glass curtain wall . sky . tree . shrub . fence . road . water . river . vehicle . sculpture . installation . pavement . street signage . fountain"
        },
        {
          "segmentation-prompts/class_id": 1,
          "segmentation-prompts/label": "Exterior urban spaces without people",
          "segmentation-prompts/prompt": "building facade . glass curtain wall . wooden facade . tree . shrub . lawn . sky . pavement . road . water . river . lantern . sculpture . installation . street lamp . signage . fence . bridge . water feature . fountain"
        },
        {
          "segmentation-prompts/class_id": 2,
          "segmentation-prompts/label": "Interior urban spaces with people",
          "segmentation-prompts/prompt": "person . shopper . crowd . retail shelf . escalator . elevator . ceiling . floor tile . glass partition . display case . door . indoor plant . wall . window . handrail . column"
        },
        {
          "segmentation-prompts/class_id": 3,
          "segmentation-prompts/label": "Interior urban spaces without people",
          "segmentation-prompts/prompt": "retail shelf . escalator . indoor corridor . ceiling . floor tile . marble floor . glass partition . display case . wall . column . indoor plant . elevator . door . window . lighting fixture . handrail"
        },
        {
          "segmentation-prompts/class_id": 4,
          "segmentation-prompts/label": "Hotel or commercial lodging spaces",
          "segmentation-prompts/prompt": "hotel bed . furniture . sofa . carpet . marble floor . tile floor . wooden floor . ceiling . bathroom . window . curtain . lamp"
        },
        {
          "segmentation-prompts/class_id": 5,
          "segmentation-prompts/label": "Private home interiors",
          "segmentation-prompts/prompt": "sofa . bed . dining table . floor . ceiling . kitchen . bookshelf . wardrobe . window . lamp . carpet . wall"
        },
        {
          "segmentation-prompts/class_id": 6,
          "segmentation-prompts/label": "Food or drink items",
          "segmentation-prompts/prompt": "food dish . meal plate . dessert . beverage cup . coffee . drink bottle . bowl . chopsticks . spoon . dining table . person . restaurant interior"
        },
        {
          "segmentation-prompts/class_id": 7,
          "segmentation-prompts/label": "Retail products and merchandise",
          "segmentation-prompts/prompt": "fashion clothing . shoes . cosmetics . product package . merchandise . retail shelf . bag . jewelry . electronics . store window . mannequin . person . floor . wall"
        },
        {
          "segmentation-prompts/class_id": 8,
          "segmentation-prompts/label": "Human-centered portrait",
          "segmentation-prompts/prompt": "person . face . group photo . building facade . sky . tree . floor . food . animal . vehicle . indoor background"
        },
        {
          "segmentation-prompts/class_id": 9,
          "segmentation-prompts/label": "Other non-spatial content",
          "segmentation-prompts/prompt": "animal . person . vehicle . advertisement poster . text . QR code . screenshot . sculpture . meme . sky . plant . signage . graphic design . logo . map . infographic . chat record"
        }
      ]
    }
  ],
  "rai:dataLimitations": [
    "Urban-ImageNet is geographically concentrated in China and should not be interpreted as a universal global urban dataset.",
    "The source is public Weibo posts, so the corpus reflects platform demographics, posting norms, commercial popularity, hashtag practices, and visibility biases rather than a representative sample of all city users.",
    "The 1K, 10K, and 100K subsets are intentionally class-balanced for benchmarking; the full 2M corpus is unbalanced and has no predefined train/validation/test split.",
    "Post-level image-text pairing is naturally loose: one post may contain up to nine images, and the post text may describe an event, emotion, route, or commercial activity rather than a literal caption of every image.",
    "Task 3 instance masks are model-generated pseudo-labels from Grounding DINO and SAM2 with quality filtering, not exhaustive human pixel annotations.",
    "The dataset is not recommended for individual identification, face recognition, account reconstruction, surveillance, social scoring, commercial profiling, law-enforcement targeting, or demographic inference about specific users."
  ],
  "rai:dataBiases": [
    "Selection bias arises because only public social-media posts with location-specific commercial-site hashtags were collected.",
    "Population bias is likely because Weibo users over-represent digitally active, urban, and younger groups compared with all city users.",
    "Geographic and cultural bias is present because all 24 cities are in China and the text is primarily original Chinese social-media language.",
    "Visual bias is present because users tend to post photogenic, socially meaningful, popular, or commercial scenes rather than mundane or private experiences.",
    "The full 2M corpus is class-imbalanced; balanced subsets are useful for model comparison but do not reflect the natural class distribution of the source platform.",
    "Segmentation pseudo-labels may inherit Grounding DINO and SAM2 biases, including missed small objects, errors on commercial interiors, and prompt vocabulary effects."
  ],
  "rai:personalSensitiveInformation": [
    "The source posts were public, but social-media images and text can still contain contextual personal information. The release therefore uses privacy-protected derivatives rather than raw data.",
    "Original usernames and account names are removed. Post ID and User ID are released only as opaque numerical identifiers after anonymization/pseudonymization and are not intended to reconstruct original accounts.",
    "Images are resized to a maximum long edge of 512 px. Faces, license plates, QR-code-like regions, and other sensitive visual regions are blurred with automated detectors followed by human spot checks.",
    "The raw high-resolution corpus, larger than 4 TB, is not publicly released and is retained securely. The public release is substantially reduced in resolution and metadata detail.",
    "The dataset does not intentionally provide health, financial, political, religious, biometric, or exact personal identity attributes. Users must not attempt re-identification or linkage to original social-media accounts."
  ],
  "rai:dataUseCases": [
    "Validated use cases include urban scene semantic classification under the 10-class HUSIC taxonomy, category-level image-text retrieval, post-level Chinese image-text retrieval with one-to-many matching, instance segmentation benchmarking, and data-scale behavior analysis across 1K/10K/100K/2M variants.",
    "The dataset is intended for non-commercial academic research in urban perception, computational urban studies, multimodal learning, image classification, cross-modal retrieval, segmentation, and reproducible benchmarking.",
    "Construct validity: the dataset is designed to represent user-generated visual and textual traces of urban commercial space perception, not the full demographic distribution of urban residents.",
    "Uses not validated include production deployment, safety-critical navigation, individual-level behavior prediction, demographic profiling, policing, surveillance, or inference about private individuals."
  ],
  "rai:dataSocialImpact": "Potential positive impacts include supporting evidence-based urban planning, helping researchers understand how people perceive and document commercial spaces, improving non-Western and Chinese-language multimodal benchmarks, and enabling public-good research for more inclusive urban design. Potential negative impacts include privacy leakage, re-identification attempts, surveillance-oriented use, account reconstruction, and commercial profiling. Mitigations include using only public posts, anonymizing identifiers, removing usernames, blurring faces and license plates, resizing images, minimizing metadata, restricting use to non-commercial academic research, and explicitly prohibiting re-identification, surveillance, face recognition, and profiling.",
  "rai:hasSyntheticData": false,
  "rai:dataCollection": "Urban-ImageNet was collected from public Sina Weibo posts using location-specific hashtags for 61 commercial sites across 24 Chinese cities from 2019-01-01 to 2025-12-31. For each public post, the pipeline retained image attachments, original Chinese post text, timestamp, city/site metadata, place tag, and lightweight engagement metadata.",
  "rai:dataCollectionType": [
    "Web Scraping",
    "User-generated content data",
    "Software Collection",
    "Manual Human Curation"
  ],
  "rai:dataCollectionRawData": "The raw corpus consisted of public Weibo image-text posts from commercial districts. The raw high-resolution corpus exceeds 4 TB and is not publicly released; only privacy-protected, resized, anonymized derivatives are distributed.",
  "rai:dataPreprocessingProtocol": [
    "Perceptual-hash near-duplicate removal, minimum-resolution filtering, automated NSFW filtering, repeated-advertisement/spam removal, metadata minimization, image resizing to max 512 px, face/license-plate/QR-sensitive-region blurring, and human spot-check review.",
    "Original usernames and direct account identifiers were removed; released Post ID and User ID fields are opaque numerical identifiers."
  ],
  "rai:dataAnnotationProtocol": "The HUSIC 10-class labels were manually annotated by trained researchers using standardized guidelines and calibration. Task 3 pseudo-labels were generated with class-specific Grounding DINO prompts followed by SAM2 mask prediction, non-maximum suppression, area filtering, and quality review.",
  "rai:dataAnnotationPlatform": [
    "Internal researcher annotation workflow",
    "Grounding DINO",
    "SAM2"
  ],
  "rai:dataAnnotationAnalysis": "The 100K benchmark set was manually labeled by three trained researchers after calibration; paper documentation reports high inter-rater agreement. Segmentation labels are marked as model-generated pseudo-labels and should be treated differently from exhaustive human pixel annotations.",
  "rai:annotationsPerItem": "Each benchmark image has one HUSIC class label. Text-image rows provide post-level metadata. Instance segmentation files contain zero or more object annotations per image, depending on model detections and filtering.",
  "rai:machineAnnotationTools": [
    "Grounding DINO",
    "SAM2",
    "pHash deduplication",
    "face/license-plate blurring detectors"
  ],
  "rai:dataReleaseMaintenancePlan": "The dataset is released as version 1.0 for NeurIPS 2026 submission. Future updates should document checksum changes, privacy-processing changes, annotation revisions, and deprecation of superseded files.",
  "prov:wasDerivedFrom": [
    {
      "@id": "https://weibo.com"
    },
    {
      "@id": "https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet"
    }
  ],
  "prov:wasGeneratedBy": [
    {
      "@type": "prov:Activity",
      "@id": "public-weibo-collection",
      "name": "Public Weibo Commercial-Site Collection",
      "prov:type": "DataCollection",
      "prov:startedAtTime": "2019-01-01",
      "prov:endedAtTime": "2025-12-31",
      "description": "Collected public posts associated with location-specific hashtags for 61 commercial sites in 24 Chinese cities."
    },
    {
      "@type": "prov:Activity",
      "@id": "privacy-preserving-preprocessing",
      "name": "Privacy-Preserving Preprocessing",
      "prov:type": "DataProcessing",
      "description": "Applied deduplication, filtering, metadata minimization, anonymization, resizing, face/license-plate/QR-sensitive-region blurring, and manual spot checks."
    },
    {
      "@type": "prov:Activity",
      "@id": "husic-manual-annotation",
      "name": "HUSIC Manual Annotation",
      "prov:type": "DataAnnotation",
      "description": "Trained researchers annotated the balanced benchmark subsets using the 10-class HUSIC taxonomy and calibration guidelines."
    },
    {
      "@type": "prov:Activity",
      "@id": "groundingdino-sam2-pseudolabeling",
      "name": "Grounding DINO + SAM2 Instance Segmentation Pseudo-Labeling",
      "prov:type": "DataAnnotation",
      "description": "Generated boxes, detected labels, confidence scores, and COCO RLE masks using class-specific prompt vocabularies, NMS, area filtering, and review."
    }
  ]
}