Sample Dataset/Sample Dataset-README.md

---
license: cc-by-nc-sa-4.0
pretty_name: Urban-ImageNet — Sample Dataset
task_categories:
  - image-classification
  - image-to-text
  - text-to-image
  - zero-shot-image-classification
  - image-segmentation
modalities:
  - image
  - text
language:
  - zh
  - en
size_categories:
  - n<1K
tags:
  - urban-perception
  - social-media
  - weibo
  - image-text-retrieval
  - instance-segmentation
  - computational-urban-studies
  - urban-ai
  - chinese-cities
  - husic
  - cross-modal-retrieval
  - multi-modal
  - scene-classification
  - urban-space-perception
  - sample-dataset
---


# 🏙️ Urban-ImageNet — Sample Dataset

**A balanced 100-image quality-inspection sample drawn from the Urban-ImageNet 100K benchmark, covering all three annotation tasks.**

<p align="center">
  <a href="https://arxiv.org/abs/2605.09936"><img src="https://img.shields.io/badge/arXiv-2605.09936-b31b1b.svg" alt="arXiv"/></a>
  <a href="https://github.com/yiasun/dataset-2"><img src="https://img.shields.io/badge/GitHub-yiasun%2Fdataset--2-black?logo=github" alt="GitHub"/></a>
  <a href="https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet"><img src="https://img.shields.io/badge/🤗%20Full%20Dataset-Urban--ImageNet-yellow" alt="Full Dataset"/></a>
</p>

> This repository is a **compact, self-contained sample** of Urban-ImageNet, designed to let anyone quickly inspect data quality, annotation completeness, and label fidelity without downloading the 100K large corpus (6+ GB). It faithfully mirrors the structure, format, and annotation conventions of the 100K benchmark across all three tasks.

---

## Overview

Urban-ImageNet is a large-scale multi-modal dataset for urban space perception (see the [full dataset page](https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet) and [paper](https://arxiv.org/abs/2605.09936) for complete documentation). This sample repository provides:

- **100 images** — 10 HUSIC classes × 10 images per class, drawn from the **training split** of the 100K dataset.
- **Complete Task 1 labels** — images stored in `ImageFolder`-compatible class subdirectories.
- **Complete Task 2 annotations** — the corresponding 100 rows from the 100K text-image pairs file, with all metadata columns preserved.
- **Complete Task 3 annotations** — a filtered COCO-format JSON containing all instance segmentation masks for the 100 images.
- **Instance mask visualisations** — side-by-side original / coloured-mask / legend images for every one of the 100 images.

This sample represents approximately **0.1%** of the 100K benchmark by image count and is strictly class-balanced, making it representative of the 100K dataset's overall structure and label distribution.

---

## How the Sample Was Created

The sample was constructed through the following transparent, reproducible procedure:

1. **Source pool:** The **training split** of the Urban-ImageNet 100K balanced benchmark (80,000 images; 8,000 per HUSIC class).

2. **Stratified random sampling:** For each of the 10 HUSIC classes, **10 images were drawn uniformly at random** without replacement, independently per class, guaranteeing strict class balance (10 images × 10 classes = 100 images total).

3. **Privacy verification:** All selected images had already undergone automated face blurring, licence-plate blurring, and QR-code blurring as part of the full dataset pipeline. The 100 selected images were additionally subject to **manual human inspection** to confirm that all privacy-sensitive regions are fully obscured before inclusion in this sample.

4. **Annotation extraction:**
   - **T1:** Images placed into their HUSIC class subdirectory — no additional annotation step required.
   - **T2:** The 100 `Image Filename` values were used to filter rows from the 100K `train.xlsx` file. All original metadata columns are preserved without modification.
   - **T3:** The 100 image filenames were used to filter entries from the 100K COCO annotation file, extracting all matching `images` entries and their associated `annotations`. Image IDs and annotation IDs are re-indexed sequentially within the sample file.

5. **Mask visualisation generation:** For each of the 100 images, a side-by-side visualisation was rendered using a standard pipeline. Each visualisation shows: left panel — original image; right panel — instance mask overlay with per-instance colour coding; right margin — colour-coded legend of detected object labels.

---

## File Structure

```
Sample Dataset/
│
├── 01 Images with labels/              ← Task 1: Scene Classification
│   ├── Exterior urban spaces with people/        (10 images)
│   ├── Exterior urban spaces without people/     (10 images)
│   ├── Food or drink items/                      (10 images)
│   ├── Hotel or commercial lodging spaces/       (10 images)
│   ├── Human-centered portrait/                  (10 images)
│   ├── Interior urban spaces with people/        (10 images)
│   ├── Interior urban spaces without people/     (10 images)
│   ├── Other non-spatial content/                (10 images)
│   ├── Private home interiors/                   (10 images)
│   └── Retail products and merchandise/          (10 images)
│                                        ────────────────────
│                                        Total: 100 images
│
├── 02 Text-Image Pairs/                ← Task 2: Cross-Modal Retrieval
│   └── Sample Dataset Paired Texts.xlsx
│
└── 03 Instance Segmentation/           ← Task 3: Instance Segmentation
    ├── Sample Dataset Annotations.json
    └── Visualization of instance labels/
        ├── Exterior urban spaces with people/    (10 visualisations)
        ├── Exterior urban spaces without people/ (10 visualisations)
        ├── Food or drink items/                  (10 visualisations)
        ├── Hotel or commercial lodging spaces/   (10 visualisations)
        ├── Human-centered portrait/              (10 visualisations)
        ├── Interior urban spaces with people/    (10 visualisations)
        ├── Interior urban spaces without people/ (10 visualisations)
        ├── Other non-spatial content/            (10 visualisations)
        ├── Private home interiors/               (10 visualisations)
        └── Retail products and merchandise/      (10 visualisations)
```

> **Image format:** All images are JPEG, privacy-protected (faces, licence plates, and QR codes blurred), resized to a maximum long edge of 512 px.

---

## Correspondence Between Files

All three annotation modalities share a common **image filename key** in the format `{UserID}_{PostTime}_{Index}` (e.g., `1197195715_2023年12月15日_5`). The table below shows how each component references this key:

| Component | Key field | Example value |
|-----------|-----------|---------------|
| `01 Images with labels/{class}/` | Filename stem (without `.jpg`) | `1197195715_2023年12月15日_5.jpg` |
| `Sample Dataset Paired Texts.xlsx` | `Image Filename` column | `1197195715_2023年12月15日_5` |
| `Sample Dataset Annotations.json` | `images[].file_name` | `1197195715_2023年12月15日_5.jpg` |
| `Visualization of instance labels/{class}/` | Filename stem (without `.jpg`) | `1197195715_2023年12月15日_5.jpg` |

---

## Task 1: Scene Classification Labels

Images are stored in an `ImageFolder`-compatible layout. The subdirectory name is the ground-truth HUSIC label, directly loadable with standard PyTorch or TensorFlow pipelines:

```python
from torchvision.datasets import ImageFolder

dataset = ImageFolder(root="Sample Dataset/01 Images with labels")
# dataset.classes → ['Exterior urban spaces with people', ...]
# dataset.class_to_idx → {'Exterior urban spaces with people': 0, ...}
print(len(dataset))  # 100
```

The 10 HUSIC classes (integer IDs assigned by `ImageFolder`'s lexicographic sort):

| Class ID | Class Label |
|---------:|-------------|
| 0 | Exterior urban spaces with people |
| 1 | Exterior urban spaces without people |
| 2 | Food or drink items |
| 3 | Hotel or commercial lodging spaces |
| 4 | Human-centered portrait |
| 5 | Interior urban spaces with people |
| 6 | Interior urban spaces without people |
| 7 | Other non-spatial content |
| 8 | Private home interiors |
| 9 | Retail products and merchandise |

> **Note:** In the full Urban-ImageNet dataset, HUSIC class IDs 0–9 follow the theoretical ordering (exterior-before-interior, spatial-before-non-spatial). When loading with `ImageFolder`, always use `class_to_idx` to map label strings to IDs rather than assuming a fixed integer correspondence.

---

## Task 2: Text-Image Pair Metadata

`Sample Dataset Paired Texts.xlsx` contains **100 rows** — one per image — with all metadata columns from the original 100K dataset preserved without modification.

### Column Schema

| Column | Type | Description |
|--------|------|-------------|
| `Image Label` | string | HUSIC class label |
| `Image Filename` | string | Join key linking to image file and T3 annotation |
| `Post ID` | integer | Anonymised numerical post identifier |
| `User ID` | integer | Anonymised numerical user identifier |
| `Post Time` | string | Original post timestamp |
| `Post Text` | string | Original Weibo post text (Chinese, unmodified) |
| `City` | string | City of the location tag |
| `Place Tag` | string | Location hashtag / commercial-site place tag |
| `Posting Tool` | string | Client or posting-source string |
| `Mentioned Users` | string | Anonymised or empty mentioned-user field |
| `Extracted Topics` | string | Topic / hashtag terms extracted from post text |
| `Extracted Locations` | string | Location mentions extracted from post text |
| `Like Count` | integer | Public engagement count at collection time |
| `Repost Count` | integer | Public repost count at collection time |
| `Comment Count` | integer | Public comment count at collection time |

> `Post Text` retains original Chinese to preserve linguistic authenticity for Task 2 evaluation. All columns retain the original data types and formatting from the 100K source file.

---

## Task 3: Instance Segmentation Annotations

`Sample Dataset Annotations.json` uses the same extended COCO format as the full Urban-ImageNet dataset.

### JSON Structure

```json
{
  "info": {
    "description": "Urban-ImageNet Instance Segmentation Annotations",
    "split": "train",
    "version": "1.0",
    "annotation_tool": "Grounding DINO + SAM2"
  },
  "categories": [
    {"id": 0, "name": "Exterior urban spaces with people"},
    ...
  ],
  "images": [
    {
      "id": 0,
      "file_name": "1197195715_2023年12月15日_5.jpg",
      "width": 512,
      "height": 384,
      "classification_label": 0
    }
  ],
  "annotations": [
    {
      "id": 0,
      "image_id": 0,
      "category_id": 0,
      "detected_label": "person",
      "detection_score": 0.8732,
      "bbox": [x, y, width, height],
      "area": 4512,
      "segmentation": {"counts": "...", "size": [384, 512]},
      "iscrowd": 0
    }
  ]
}
```

**Extended fields beyond standard COCO:**

- `classification_label` *(in `images`)*: HUSIC class ID — enables multi-task joint training.
- `detected_label` *(in `annotations`)*: the specific object term detected by Grounding DINO (e.g., `"escalator"`, `"retail shelf"`, `"hotel bed"`).
- `detection_score` *(in `annotations`)*: Grounding DINO confidence score, enabling threshold-based downstream filtering.
- Segmentation masks are stored in **COCO RLE format**, directly compatible with `pycocotools`.

### Instance Mask Visualisations

The `Visualization of instance labels/` subfolder provides a visual verification of every annotation in this sample. Each visualisation image shows:

- **Left half:** the original image.
- **Right half:** the same image with coloured per-instance mask overlays (each unique instance receives a distinct colour).
- **Right margin:** a colour-coded legend listing the `detected_label` for each coloured mask.

Visualisation filenames match their source images exactly (e.g., `1197195715_2023年12月15日_5.jpg`), within the same class subdirectory, enabling direct one-to-one correspondence without any filename transformation.

---

## Relationship to the Full Dataset

| Property | This Sample | 100K Benchmark |
|----------|-------------|----------------|
| Images | 100 | 100,000 |
| Classes | 10 | 10 |
| Images per class | 10 | 10,000 |
| Source split | train only | train / val / test |
| T1 labels | ✅ | ✅ |
| T2 text metadata | ✅ 100 rows | ✅ 80K / 10K / 10K rows |
| T3 instance annotations | ✅ filtered JSON | ✅ train / val / test JSON |
| T3 visualisations | ✅ all 100 images | ✅ selected examples |
| Privacy protection | ✅ automated + manual review | ✅ automated + manual review |
| File format & schema | Identical to full dataset | — |

Code written against this sample runs unchanged on the full dataset by substituting the root path.

---

## Quick-Start

```python
import json
import pandas as pd
from pathlib import Path
from torchvision.datasets import ImageFolder

root = Path("Sample Dataset")

# ── Task 1 ──────────────────────────────────────────────────────────────
dataset = ImageFolder(root=root / "01 Images with labels")
print(dataset.classes)   # 10 HUSIC class names
print(len(dataset))      # 100

# ── Task 2 ──────────────────────────────────────────────────────────────
df = pd.read_excel(root / "02 Text-Image Pairs" / "Sample Dataset Paired Texts.xlsx")
print(df.shape)                 # (100, 15)
print(df["Post Text"].iloc[0])  # Chinese post text

# ── Task 3 ──────────────────────────────────────────────────────────────
with open(root / "03 Instance Segmentation" / "Sample Dataset Annotations.json") as f:
    coco = json.load(f)
print(f"Images:      {len(coco['images'])}")       # 100
print(f"Annotations: {len(coco['annotations'])}")  # typically 500–1500

# Join image filename → annotations
img_id_map = {img["file_name"]: img["id"] for img in coco["images"]}
ann_by_img = {}
for ann in coco["annotations"]:
    ann_by_img.setdefault(ann["image_id"], []).append(ann)
```

---

## Privacy and Responsible Use

All images are derived from **public Weibo posts** and have undergone:

- Automated face detection and blurring (all detected faces).
- Automated licence-plate detection and blurring.
- Automated QR-code detection and blurring, supplemented by manual spot-checks.
- **Additional manual human review** of all 100 sample images, confirming complete protection of all privacy-sensitive regions.

Original usernames have been replaced with opaque numerical identifiers. Images are released at ≤ 512 px long edge. Use is restricted to **non-commercial academic research**. Re-identification, biometric profiling, facial recognition, and surveillance applications are strictly prohibited. See the [full dataset page](https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet) for the complete data-use agreement.

---

## Citation

```bibtex
@article{ou2026urbanimagenet,
  title   = {Urban-ImageNet: A Large-Scale Multi-Modal Dataset and Evaluation Framework for Urban Space Perception},
  author  = {Ou, Yiwei and Cheung, Chung Ching and Ang, Jun Yang and Ren, Xiaobin and Sun, Ronggui and Gao, Guansong and Zhao, Kaiqi and Manfredini, Manfredo},
  journal = {arXiv preprint arXiv:2605.09936},
  year    = {2026},
  eprint  = {2605.09936},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url     = {https://arxiv.org/abs/2605.09936}
}
```

**Paper:** [arXiv:2605.09936](https://arxiv.org/abs/2605.09936)  
**Full Dataset:** [huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet](https://huggingface.co/datasets/Yiwei-Ou/Urban-ImageNet)  
**Benchmark code:** [github.com/yiasun/dataset-2](https://github.com/yiasun/dataset-2)

---

## License

Released under **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)**.  
Non-commercial academic research use only. See [LICENSE](https://creativecommons.org/licenses/by-nc-sa/4.0/) for full terms.