README.md

---
annotations_creators:
- expert-generated
language:
- en
- zh
language_creators:
- found
- expert-generated
license: cc-by-4.0
multilinguality:
- multilingual
pretty_name: 'MMS-VPR: Multimodal Street-Level Visual Place Recognition Dataset'
size_categories:
- 100K<n<1M
source_datasets:
- original
task_categories:
- image-classification
- video-classification
- image-to-text
- zero-shot-classification
task_ids:
- multi-class-image-classification
- multi-label-classification
tags:
- visual-place-recognition
- multimodal-learning
- graph-neural-networks
- urban-computing
- pedestrian-navigation
- day-night-recognition
- temporal-analysis
- spatial-graph
- geospatial
- computer-vision
paperswithcode_id: mms-vpr
---

# MMS-VPR: Multimodal Street-Level Visual Place Recognition Dataset

[![Paper](https://img.shields.io/badge/Paper-arXiv-red)](https://arxiv.org/abs/2505.12254)
[![Dataset](https://img.shields.io/badge/Dataset-HuggingFace-yellow)](https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR)
[![Benchmark](https://img.shields.io/badge/Benchmark-GitHub-blue)](https://github.com/yiasun/MMS-VPRlib)
[![License: CC BY 4.0](https://img.shields.io/badge/License-CC%20BY%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by/4.0/)

## Overview

**MMS-VPR** is the first large-scale multimodal street-level visual place recognition dataset featuring comprehensive integration of **images**, **videos**, and **rich textual annotations** with **day-night coverage** and **7-year temporal span** in dense pedestrian-only environments.

**Keywords**: Visual Place Recognition, Multimodal Learning, Pedestrian Navigation, Urban Computing, Graph Neural Networks, Day-Night VPR, Street-Level Localization

### Key Features

- 🚶 **Pedestrian-Only Perspective**: First VPR dataset systematically collected in dense pedestrian commercial districts
- 🌓 **Day-Night Coverage**: Balanced temporal sampling across daytime (7AM-5PM) and nighttime (6PM-10PM)
- 🎯 **Multimodal**: Images + Videos + Text annotations (GPS, store names, spatial metrics)
- 📅 **7-Year Temporal Span**: Field collection (2024) + Social media data (2019-2025)
- 🗺️ **Graph Structure**: 208 locations organized in spatial graph with connectivity relationships
- 🏙️ **Urban Science Integration**: Space syntax metrics enriching spatial configuration context

### Dataset Statistics

| Modality | Count/Coverage | Source | Details |
|----------|----------------|--------|---------|
| **Images** | 110,529 images | Field + Social Media | Resolution: 256×192 (preprocessed) |
| **Videos** | 2,527 clips | Field Collection | 20-60s, 256×144, 30fps |
| **Texts** | 208 locations | OCR + Manual | GPS, store names, signage, spatial metrics |
| **Graph Structure** | 208 nodes/edges | Urban Network | Connectivity, distances, space syntax |
| **Locations** | 208 unique places | Chengdu Taikoo Li | ~70,800 m² pedestrian district |
| **Temporal Coverage** | 7 years | 2019-2025 | Fine-grained + long-term span |

**Spatial Composition:**
- 81 Nodes (street intersections)
- 125 Edges (61 horizontal + 64 vertical streets)  
- 2 Squares (large open spaces)

---

## Visual Overview

### Dataset Framework

Our dataset addresses four critical limitations in existing VPR datasets through systematic multimodal data collection:

![MMS-VPR Framework](01-Overall-Framework.png)

*Figure 1: MMS-VPR framework addressing four VPR limitations: (1) pedestrian-only perspective, (2) day-night coverage, (3) multimodal integration (images + videos + text), and (4) 7-year temporal span enhanced with social media data.*

### Data Collection Pipeline

![Data Collection](02-Data-Collection-Methods.png)

*Figure 2: Systematic methodology—site collection → data processing → textual annotation → social media integration*

### Benchmark Platform

![Benchmark Workflow](03-MMS-VPRlib-Benchmark.png)

*Figure 3: MMS-VPRlib unified platform integrating datasets, models, and evaluation pipelines. [GitHub Repository](https://github.com/yiasun/MMS-VPRlib)*

---

## Dataset Structure

The dataset comprises four main components:

```
MMS-VPR/
├── Images/                        # 110,529 images across 208 location folders
│   ├── N-1-1/                     # Node (intersection) images
│   ├── Eh-1-1/                    # Horizontal edge (street) images
│   ├── Ev-1-1/                    # Vertical edge (street) images
│   └── S-1/                       # Square images
├── Videos/                        # 2,527 video clips
│   ├── N-1-1/                     # Videos organized by location
│   └── ...
├── Texts/                         # Textual annotations and metadata
│   ├── Annotations.xlsx           # Location labels, GPS, store names, signage
│   ├── Metadata-Images.xlsx       # Image EXIF metadata
│   └── Metadata-Videos.xlsx       # Video metadata
└── Graph Structure/               # Spatial graph organization
    ├── Graph_Structure_README.md  # Detailed graph documentation
    ├── 00 Street Network Graph.pdf
    ├── 01 Node Features.xlsx
    ├── 02 Edge Features.xlsx
    ├── 03 Edge Connections.xlsx
    └── 04 Square Features.xlsx
```

### Location Encoding System

Each location follows a hierarchical encoding scheme:

- **Nodes** (Intersections): `N-i-j` where i=row, j=column (e.g., `N-1-1`, `N-2-3`)
- **Horizontal Edges** (Streets): `Eh-i-j` for east-west streets (e.g., `Eh-1-1`)
- **Vertical Edges** (Streets): `Ev-j-i` for north-south streets (e.g., `Ev-1-1`)
- **Squares**: `S-k` ranked by area (e.g., `S-1` is the largest)

This encoding preserves both geometric position and topological relationships, enabling graph-based learning.

## Multimodal Data Composition

### 1. Image Data (110,529 images)

**Field Collection** (78,575 images):
- Systematic coverage from 4 cardinal directions (N, S, E, W)
- Dual perspectives: horizontal (0°) + upward (45°)
- Day-night balanced sampling (7AM-10PM)
- 1Hz frame extraction from videos + standalone photos

**Social Media** (31,954 images):
- Curated from Weibo (2019-2025)
- Georeferenced to 208 locations
- Diverse viewpoints and 7-year temporal extent

**Resolution**: 256×192 (preprocessed), original 4032×3024 available upon request

### 2. Video Data (2,527 clips)

- Duration: 20-60 seconds per clip
- Resolution: 256×144 @ 30fps (preprocessed), original 1920×1080
- Captured along streets from multiple directions
- Day-night coverage enabling motion-aware place recognition

### 3. Textual Annotations

Each location includes:

**Spatial Identifiers:**
- Systematic codes preserving graph topology (e.g., `N-3-4`, `Eh-5-2`)
- Enable adjacency matrix construction for GNN applications

**Semantic Text:**
- Store names and visible signage (OCR-extracted + manually verified)
- Example: "Starbucks, Adidas, LEGO, MUJI"
- Supports text-based retrieval and multimodal fusion

**Geospatial Data:**
- GPS coordinates (longitude, latitude)
- Altitude information from EXIF metadata

**Space Syntax Metrics** (from urban science):
- Integration: Global accessibility measure
- Betweenness: Through-movement potential
- Computed using both angular and weighted distances
- See `Graph_Structure_README.md` for details

## Graph-Based Organization

All 208 locations form a spatial graph **G = (V, E)** representing the pedestrian network:

- **Nodes (V)**: 81 street intersections + 2 squares
- **Edges (E)**: 125 pedestrian street segments
- **Attributes**: Physical properties, connectivity, turning angles, distances

The `Graph Structure/` folder provides complete:
- Network visualization (PDF)
- Node/edge feature tables
- Connection tables with angular/Euclidean distances
- Ready-to-use adjacency matrices for GNN research

**For detailed graph documentation, see `Graph_Structure_README.md`**

---

## Download and Usage

> **Note on Dataset Viewer**: This dataset is distributed as compressed archives (.tar.gz) for efficient storage and distribution. The Hugging Face Dataset Viewer is not available for this format, which is standard for large-scale image datasets (similar to ImageNet, Places365). This does not affect dataset usage.

### Quick Start

The dataset is split into four components for flexible downloading (~11GB total):

```bash
# Download all components
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Images.tar.gz
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Videos.tar.gz
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Texts.tar.gz
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Graph_Structure.tar.gz

# Extract all
tar -xzf Images.tar.gz
tar -xzf Videos.tar.gz
tar -xzf Texts.tar.gz
tar -xzf Graph_Structure.tar.gz

# Dataset is ready to use!
```

### Selective Download (Save Bandwidth)

Download only what you need:

**For image-only experiments** (~2.3GB):
```bash
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Images.tar.gz
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Texts.tar.gz
tar -xzf Images.tar.gz
tar -xzf Texts.tar.gz
```

**For video-based research** (~8.8GB):
```bash
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Videos.tar.gz
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Texts.tar.gz
tar -xzf Videos.tar.gz
tar -xzf Texts.tar.gz
```

**For graph-based methods** (minimal):
```bash
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Graph_Structure.tar.gz
wget https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR/resolve/main/Texts.tar.gz
tar -xzf Graph_Structure.tar.gz
tar -xzf Texts.tar.gz
```

### File Breakdown

| Component | Size | Contents | Use Case |
|-----------|------|----------|----------|
| `Images.tar.gz` | 2.25 GB | 110,529 images (256×192) | Image-based VPR |
| `Videos.tar.gz` | 8.78 GB | 2,527 videos (256×144) | Video-based VPR, motion analysis |
| `Texts.tar.gz` | 417 KB | Annotations, metadata | Multimodal learning, text features |
| `Graph_Structure.tar.gz` | 112 KB | Spatial graph, topology | GNN-based methods, spatial analysis |
| **Total** | **~11 GB** | Complete multimodal dataset | All experiments |

### Using Hugging Face Hub (Alternative)

```python
from huggingface_hub import hf_hub_download

# Download specific components
images_path = hf_hub_download(
    repo_id="Yiwei-Ou/MMS-VPR",
    filename="Images.tar.gz",
    repo_type="dataset"
)

texts_path = hf_hub_download(
    repo_id="Yiwei-Ou/MMS-VPR",
    filename="Texts.tar.gz",
    repo_type="dataset"
)
```

### After Extraction

Your directory structure will be:
```
MMS-VPR/
├── Images/           # 208 location folders with images
├── Videos/           # 208 location folders with video clips
├── Texts/            # Annotation files (xlsx format)
└── Graph Structure/  # Network topology and features
```

### Original High-Resolution Data

The preprocessed dataset uses reduced resolution for efficiency. Original high-resolution data (~120GB+) is available upon reasonable request:
- Images: 4032×3024 (original camera resolution)
- Videos: 1920×1080 @ 30fps

Please contact the authors via GitHub issues or email for access to original data.

---

## Data Collection Methodology

### Site Selection

**Location**: Chengdu Taikoo Li, Chengdu, China
- Area: ~70,800 m²
- Type: Open-air commercial district (pedestrian-only)
- Characteristics: Dense retail, dining, leisure, cultural spaces

### Collection Principles

Our methodology addresses three critical VPR challenges:

**1. Four-Direction Coverage**
- Systematic capture from N, S, E, W directions
- Addresses viewpoint variation (40% performance drop when viewpoints differ)
- Superior to 360° panoramas (avoid geometric distortions)

**2. Dual-Perspective Capture**
- Horizontal (0°): Eye-level navigation features
- Upward (45°): Building facades and landmarks
- Matches human visual field strategies in urban environments

**3. Balanced Day-Night Sampling**
- Daytime: 7AM-5PM
- Nighttime: 6PM-10PM
- Equal data volume for illumination-robust learning

### Equipment

- **Consumer smartphones**: iPhone XS Max, iPhone 11 Pro Max
- **Accessible and reproducible**: No specialized equipment required
- Framework designed for replication in other cities

### Pipeline

1. **Site Collection**: Systematic field surveys (2024)
2. **Data Processing**: Resolution standardization, frame extraction, cleaning
3. **Textual Annotation**: GPS, store names, OCR signage, spatial metrics
4. **Social Media Integration**: Weibo data (2019-2025) for temporal span

---

## Use Cases

### Visual Place Recognition
- Street-level localization in pedestrian environments
- Cross-temporal place matching (day-night, seasonal)
- Viewpoint-invariant recognition

### Multimodal Learning
- Vision-language models for place recognition
- Video-based motion-aware VPR
- Text-guided visual retrieval

### Graph Neural Networks
- Spatial relationship modeling
- GNN-based place recognition
- Graph-constrained retrieval

### Urban Analytics
- Pedestrian flow analysis
- Spatial accessibility studies
- Commercial district characterization

### Augmented Reality & Navigation
- Pedestrian AR applications
- Indoor-outdoor transition scenarios
- Context-aware wayfinding

---

## Benchmark Platform

**MMS-VPRlib**: Unified benchmarking platform consolidating:
- Multiple datasets (Pittsburgh, Tokyo 24/7, Nordland, MMS-VPR)
- 17+ baseline models (CNN, RNN, Transformer, multimodal)
- Standardized evaluation pipeline
- Modular components for data processing, modeling, fusion

**Repository**: [https://github.com/yiasun/MMS-VPRlib](https://github.com/yiasun/MMS-VPRlib)

---

## Privacy and Ethics

All data collection adheres to ACM and KDD ethical guidelines:

### Privacy Protection
- ✅ All faces and license plates automatically detected and pixelated
- ✅ Manual review to verify PII removal
- ✅ Public outdoor spaces only (no private/restricted areas)
- ✅ Preprocessed low-resolution version (256×144) further protects privacy

### Ethical Data Collection
- ✅ Pedestrian-level viewpoints (natural eye-level, not surveillance)
- ✅ Public accessible areas during normal hours (7AM-10PM)
- ✅ No intrusion during late-night hours
- ✅ Systematic, unbiased spatial and temporal coverage

**Note**: While automated anonymization has been applied, if you identify any privacy concerns, please contact us immediately.

---

## License

This dataset is released under [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).

You are free to:
- **Share**: Copy and redistribute the material
- **Adapt**: Remix, transform, and build upon the material
- **Commercial use**: Use for any purpose, including commercially

Under the following terms:
- **Attribution**: You must give appropriate credit and indicate if changes were made

---

## Acknowledgments

We thank all contributors to this dataset. Field data were collected by the research team in 2024. Social media data were curated from publicly shared posts on Weibo (2019-2025).

---

## Citation

If you use this dataset in your research, please cite:

```bibtex
@misc{ou2025mmsvpr,
  title        = {MMS-VPR: Multimodal Street-Level Visual Place Recognition Dataset and Benchmark},
  author       = {Ou, Yiwei and Ren, Xiaobin and Sun, Ronggui and Gao, Guansong and Zhao, Kaiqi and Manfredini, Manfredo},
  year         = {2025},
  eprint       = {2505.12254},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2505.12254}
}
```

---

## Contact

For questions, issues, or requests for original high-resolution data:

- **GitHub Issues**: [MMS-VPR Repository](https://github.com/yiasun/MMS-VPRlib/issues)
- **Email**: you661@aucklanduni.ac.nz
- **Hugging Face**: [Dataset page](https://huggingface.co/datasets/Yiwei-Ou/MMS-VPR)

---

## Updates

**Latest Version**: v2.0 (February 2026)
- 208 locations (updated from 207)
- Integrated social media data (2019-2025)
- Enhanced textual annotations
- Complete graph structure documentation