MMS-VPR: A Fine-Grained Multimodal Street-Level Visual Place Recognition Dataset and Evaluation Benchmark for Dense Pedestrian Environments

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 a 7-year temporal span in dense pedestrian-only environments.

MMS-VPR comprises 110,529 images and 2,527 video clips across 208 fine-grained location classes within a ~70,800 m² compact urban district in Chengdu, China — with adjacent classes separated by as little as 10–20 m. Unlike large-area VPR datasets built on vehicle-mounted imagery, MMS-VPR elevates recognition difficulty through visual ambiguity at the pedestrian scale: repetitive storefronts, dense crowd occlusion, rapid viewpoint changes, and dramatic day–night appearance shifts that challenge both CNN-based and Transformer-based state-of-the-art methods.

We also release MMS-VPRlib, a unified benchmarking platform consolidating 22 baselines spanning shallow ML, CNN, Transformer, GNN, and vision–language architectures under a standardized, reproducible evaluation pipeline across 7 VPR datasets.

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

Key Features

• 🚶 Pedestrian-Only Perspective: VPR dataset systematically collected in dense urban pedestrian districts inaccessible to vehicles
• 🌓 Day–Night Coverage: Balanced temporal sampling across daytime (7AM–5PM) and nighttime (6PM–10PM)
• 🎯 Multimodal: Images + Videos + Text annotations (GPS, store names/signage, spatial identifiers, space syntax metrics)
• 📅 7-Year Temporal Span: Field collection (2024) + Social media data (2019–2025) at the same 208 locations
• 🗺️ Spatial Graph Structure: 208 locations organized in an explicit pedestrian network graph with full connectivity tables
• 🏙️ Urban Science Integration: Space syntax metrics (integration, betweenness) enriching spatial configuration context

---

Visual Overview

Dataset Framework

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

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.

Dataset Statistics

Modality	Count / Coverage	Source	Details
Images	110,529 images	Field (2024) + Social Media (2019–2025)	Resolution: 256×192 (preprocessed)
Videos	2,527 clips	Field Collection (2024)	20–60 s, 256×144, 30 fps
Texts	208 locations	OCR + Manual verification	GPS, store names, signage, spatial metrics
Graph Structure	208 nodes/edges	Urban pedestrian network	Connectivity, distances, space syntax
Location Classes	208 unique places	Chengdu Taikoo Li	~70,800 m² pedestrian district
Temporal Coverage	7 years	2019–2025	Fine-grained + long-term span

Spatial Composition:

Spatial Type	Code Format	Count	Images	Videos
Nodes (street intersections)	N-i-j	81	47,049	1,197
Squares (large open spaces)	S-k	2	5,686	7
Horizontal Edges (east–west streets)	Eh-i-j	61	28,178	662
Vertical Edges (north–south streets)	Ev-j-i	64	29,616	661
Total	—	208	110,529	2,527

Data Collection Pipeline

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

Benchmark Platform

Figure 3: MMS-VPRlib unified platform integrating datasets, models, and evaluation pipelines. GitHub Repository

---

Comparison with Existing Datasets

Dataset	Images	Videos	Text	Day–Night	Multi-view	Long-term	Pedestrian-only	Graph labels
Tokyo 24/7	✓	✗	✗	✓	◑	◑	◑	✗
Pittsburgh	✓	✗	✗	✗	✓	◑	✗	✗
Nordland	✓	✓	✗	◑	✗	✓	✗	✗
Cambridge	✓	✗	✗	✓	✓	◑	◑	✗
GSV-Cities	✓	✗	✗	◑	✓	✓	✗	✗
MMS-VPR (ours)	✓	✓	✓	✓	✓	✓	✓	✓

✓ = supported; ◑ = partial; ✗ = not provided.

The key distinction of MMS-VPR is not geographic breadth alone, but the combination of pedestrian-only access, fine spatial granularity (~10–20 m class separation), multimodality, balanced day–night sampling, and long-term temporal augmentation at the same 208 locations.

---

Sample Dataset

A curated sample (38MB) of MMS-VPR is available for rapid data quality inspection without downloading the full dataset (~11 GB):

Sample dataset

The sample covers all four spatial types (east–west street, north–south street, intersection node, open square) with images across multiple perspectives, day/night conditions, weather variations, and a social media temporal enhancement dataset spanning 2019–2025. The 03 Texts and 04 Graph folders in the sample are complete and identical to the full dataset; only 01 Images and 02 Videos have been manually subsampled:

Location	Type	Sample Content
Eh-12-7	Horizontal Edge	8 images (4 directions × day/night) + 12 video clips
Ev-1-8	Vertical Edge	8 images (multi-direction × sunny/rain/night)
N-14-1	Node	16 images (multi-perspective × day/night/rain)
S-1	Square	28 social media images across 2019–2025 (28 distinct events)

Sampling Methodology

Because the full dataset contains hundreds to thousands of highly similar frames per location at fine spatial granularity, conventional random sampling would likely oversample a single direction or lighting condition, misrepresenting the dataset's true breadth. We therefore used manual expert selection to guarantee coverage of all spatial types, viewing directions, lighting conditions, weather variants, and the social media temporal modality — while using the fewest possible files. Every file in the sample is drawn directly from the corresponding location in the full dataset without modification.

See Sample dataset/Sample Dataset-README.md for complete folder structure documentation and sampling details.

---

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 (east–west street) images
│   ├── Ev-1-1/                    	# Vertical edge (north–south street) images
│   └── S-1/                       	# Square images
├── Videos/                        	# 2,527 video clips
│   ├── N-1-1/                     	# Videos organized by location
│   └── ...
├── Texts/                         	# Textual annotations and metadata
│   ├── Textual Annotations.xlsx   	# Location labels, GPS, store names, signage
│   ├── Metadata-Images.xlsx       	# Image EXIF metadata
│   └── Metadata-Videos.xlsx       	# Video metadata
└── Graph Structure/               	# Spatial graph data
│   ├── 00 Street Network Graph.pdf
│   ├── 01 Node Features.xlsx
│   ├── 02 Edge Features.xlsx
│   ├── 03 Edge Connections.xlsx
│   ├── 04 Square Features.xlsx
│   └── Graph_Structure_README.md 	# Detailed graph documentation
└── Sample dataset/         		# Representative sample dataset

Location Encoding System

Each location follows a hierarchical encoding scheme that preserves both geometric position and topological relationships:

• Nodes (Intersections): N-i-j — row i, column j (e.g., N-1-1, N-2-3)
• Horizontal Edges (East–West Streets): Eh-i-j (e.g., Eh-1-1, Eh-5-3)
• Vertical Edges (North–South Streets): Ev-j-i (e.g., Ev-1-1, Ev-2-4)
• Squares: S-k ranked by area (e.g., S-1 is the largest open square)

This encoding enables direct adjacency matrix construction for GNN-based methods. Full encoding methodology and connectivity tables are in Graph Structure/Graph_Structure_README.md.

---

Multimodal Data Composition

1. Image Data (110,529 images)

Field Collection (2024) — 78,575 images:

• Systematic coverage from 4 cardinal directions (N, S, E, W) for each location
• Dual perspectives: horizontal (0°, eye-level) + upward (45°, facade/landmark)
• Balanced day–night sampling: daytime (7AM–5PM) and nighttime (6PM–10PM)
• Multiple weather conditions: sunny, overcast, rain
• 1 Hz frame extraction from 2,527 raw video clips + standalone photographs
• Raw field data: >90 GB; preprocessed to 256×192

Social Media (2019–2025) — 31,954 images:

• Curated from Weibo (Chinese social media platform), keyword-matched and geolocated
• Spans January 2019 through December 2025 across all 208 locations
• Diverse viewpoints, event-driven appearance changes (seasonal installations, commercial turnover)
• Provides long-term temporal modeling capability absent in field-only datasets

Resolution: 256×192 (preprocessed).

2. Video Data (2,527 clips)

• Duration: 20–60 seconds per clip
• Resolution: 256×144 @ 30 fps (preprocessed); original 1920×1080
• Captured along all 208 locations from multiple directions
• Comprehensive day–night coverage enabling motion-aware temporal place recognition
• Filmed using a handheld gimbal for stabilization

3. Textual Annotations

Each of the 208 locations is annotated with:

Spatial Identifiers:

• Systematic codes encoding graph topology (e.g., N-3-4, Eh-5-2, Ev-1-8)

• Enable direct adjacency matrix construction for GNN applications

Semantic Signage Text:

• Store names and visible signage extracted via OCR and manually verified

• Cleaned and de-duplicated to retain discriminative identifiers per location

• Examples: Starbucks, Adidas, LEGO, MUJI, WM HOUSE

• Supports text-based retrieval and multimodal vision–language fusion

Geospatial Data:

• GPS coordinates (longitude, latitude, altitude) from EXIF metadata

Space Syntax Metrics (from urban science — provided as auxiliary enrichment for future research):

• Integration: Global accessibility measure (higher = more centrally located, more easily accessible)

• Betweenness: Through-movement potential (higher = primary pedestrian route, heavier predicted pedestrian flow)

• Computed using both angular distance and weighted distance formulations

• See Graph Structure/Graph_Structure_README.md for computation details and visualizations

4. Graph-Based Spatial Structure

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 as a zoomable vector PDF
• Node/edge feature tables (spatial, semantic, topological attributes)
• Full connection table: shared nodes, turning angles, Euclidean and angular distances between all adjacent locations
• Ready-to-use inputs for GNN-based place recognition methods (GCN, GAT, HGNN)

For complete graph documentation, see Graph Structure/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 — Download Everything (~11 GB)

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

# 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):

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):

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):

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	Primary Use Case
Images.tar.gz	~2 GB	110,529 images (256×192)	Image-based VPR
Videos.tar.gz	~9 GB	2,527 video clips (256×144)	Video-based VPR, motion analysis
Texts.tar.gz	~400 KB	Annotations + EXIF metadata	Multimodal fusion, text features
Graph_Structure.tar.gz	~100 KB	Spatial graph, topology	GNN-based methods, spatial analysis
Total	~11 GB	Complete multimodal dataset	All experiments

Using Hugging Face Hub (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

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 efficient distribution. Original high-resolution data (~120 GB+) may be available upon reasonable request:

• Images: 4032×3024 (original camera resolution)
• Videos: 1920×1080 @ 30 fps

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

---

Data Collection Methodology

Site Selection

Location: Chengdu Taikoo Li, Jinjiang District, Chengdu, China

• Area: ~70,800 m²;
• Type: Open-air urban commercial district (pedestrian-only);
• Characteristics: Dense retail, dining, leisure, art, and cultural spaces; 40+ individual 2–3-storey buildings interconnected by open-air pedestrian pathways integrated with surrounding urban streets;
• Site selection rationale: Spatial openness with natural lighting and high crowd density; functional diversity ensuring rich spatial semantics; seamless integration with surrounding urban public streets, making it representative of globally common urban pedestrianized districts.

Crucially, this environment challenges place recognition with repetitive visual elements (franchise storefronts, identical signage across locations) and dense occlusion patterns that indoor mall datasets and large-area outdoor datasets both fail to capture.

Three Collection Principles

Our methodology is grounded in three evidence-based principles:

Principle 1 — Four-Direction Coverage

VPR performance degrades by up to 40% under viewpoint mismatch. We systematically capture all four cardinal directions (N, S, E, W) for each street segment, providing coverage matching real-world social media content patterns while avoiding the geometric distortion of 360° panoramas.

Principle 2 — Dual-Perspective Capture

Pedestrians employ two natural viewing strategies in high-rise environments: forward-looking (0°) for navigation and upward-looking (30–60°) for landmark recognition. We capture each view at both horizontal (0°) and upward (45°) angles. Upward-facing imagery improves place recall by 15–30% where building facades are the primary discriminative feature.

Principle 3 — Balanced Day–Night Coverage

Illumination variation causes 50–80% performance drops when daytime-trained models are deployed at night. We enforce equivalent data volume during daytime (7AM–5PM) and nighttime (6PM–10PM), with three fine-grained daytime intervals: early morning (7–9 AM, minimal occlusion), noon (12–2 PM, maximum crowd density), and twilight (5–6 PM, mixed natural and artificial lighting).

Equipment

• Devices: iPhone XS Max, iPhone 11 Pro Max (consumer smartphones; no specialized equipment)
• Stabilization: Handheld gimbal for video capture
• Accessibility: The entire collection framework is replicable in any city using standard consumer devices

Pipeline

1. Field Collection (2024): Systematic on-site collection yielded 78,575 images and 2,527 video clips across all 208 locations, with fine-grained temporal coverage from 7AM to 10PM and varied pedestrian occlusion patterns.

2. Data Processing: Resolution standardization, frame extraction, cleaning

3. Textual Annotation: GPS, store names, OCR signage, spatial metrics

4. Social Media Integration (2019–2025): 31,954 georeferenced images from Weibo, spanning January 2019 through December 2025, enabling modeling of seasonal variation, architectural change, commercial turnover, and event-driven appearance change over seven years. Social media imagery shows a distribution across all four spatial types, with intersection nodes attracting the highest volume of publicly shared photographs. A dip in 2022 reflects COVID-19-related restrictions on public gatherings in China.

---

Benchmark Platform: MMS-VPRlib

MMS-VPRlib is an open-source unified benchmarking platform providing standardized, reproducible pipelines for multimodal VPR.

Repository: https://github.com/yiasun/MMS-VPRlib

Supported Datasets (7 total)

Dataset	Setting	Key Challenge
MMS-VPR (ours)	Pedestrian fine-grained	208 classes, sub-20 m separation
Pittsburgh	Street-view retrieval	Large-scale urban
Tokyo 24/7	City landmark	Day–night, multi-view
Nordland	Seasonal route	Seasonal appearance change
Cambridge	Landmark pose	Outdoor localization
New College	Campus route	Appearance change
CityPlace	City-scale VPR	Large-scale scenes

Supported Models (22 baselines)

Category	Models
Shallow ML	LR, GNB, SVC, KNN, RF, MLP
CNN	ResNet, PatchNetVLAD
Transformer / VPR-specialized	ViT, R2Former, BoQ, MixVPR, SALAD, SFRS, EigenPlaces, CosPlace
Graph Neural Networks	GCN, GAT, HGNN, HGNN+ResNet (graph–visual fusion)
Vision–Language	CLIP, BLIP

All models are implemented in self-contained Jupyter notebooks with conda environment files. All experiments are reproducible on a single consumer GPU (NVIDIA RTX 3060/4060, 8 GB VRAM).

---

Benchmark Results

Task Formulation

MMS-VPR is formulated as closed-set classification over C = 208 location classes. Given a query (image, video, text, or any combination), a model predicts class \hat{c} \in {0, \ldots, 207}. A prediction is correct if and only if it matches the ground-truth location label. Evaluation uses an 80/20 per-location stratified train/test split ensuring every class appears in both sets. All results are mean over 5 independent runs.

RQ1 — 22 Baseline Results on MMS-VPR (Image-Only Input, 208 Classes)

Results are means over 5 independent runs, 80/20 stratified per-location train/test split.

Note: †Pure GNN methods (GCN, GAT, HGNN) receive only graph-structural inputs (no visual features), serving as structural-only lower bounds. HGNN+ResNet denotes graph–visual fusion.

Metric	LR	GNB	SVC	KNN	RF	MLP	GCN†	GAT†	HGNN†	H+R	ResNet	PNet	ViT	R2F	BoQ	SFRS	MixVPR	SALAD	EigPl	CosPl	BLIP	CLIP
Acc	0.110	0.082	0.080	0.422	0.198	0.202	0.076	0.078	0.078	0.879	0.856	0.846	0.596	0.771	0.744	0.734	0.789	0.870	0.918	0.933	0.674	0.885
P	0.090	0.090	0.040	0.508	0.396	0.184	0.048	0.046	0.024	0.879	0.849	0.838	0.584	0.775	0.720	0.722	0.793	0.870	0.910	0.925	0.678	0.886
R	0.060	0.100	0.030	0.402	0.154	0.184	0.034	0.033	0.032	0.872	0.837	0.860	0.564	0.770	0.694	0.673	0.785	0.865	0.904	0.921	0.683	0.878
F1	0.054	0.084	0.020	0.406	0.190	0.174	0.028	0.024	0.020	0.874	0.841	0.845	0.566	0.765	0.703	0.689	0.781	0.861	0.906	0.924	0.671	0.877

Abbreviations: PNet = PatchNetVLAD; R2F = R2Former; EigPl = EigenPlaces; CosPl = CosPlace; H+R = HGNN+ResNet.

Key Findings

• VPR-specialized CosPlace achieves the strongest result (Acc: 0.933, F1: 0.924), outperforming the strongest generic backbone ResNet by +9.0%, confirming that VPR-oriented metric learning and aggregation provide gains beyond generic visual feature extraction
• EigenPlaces ranks second (Acc: 0.918), followed by SALAD (0.870) and CLIP (0.885)
• Large-scale pretraining substantially helps: CLIP improves vanilla ViT by +48.5% (0.596 → 0.885)
• HGNN+ResNet (0.879) outperforms standalone ResNet (0.856), confirming that graph-structured spatial context provides complementary signal to visual features
• Even the best model (CosPlace, 0.933) leaves 6.7% unresolved — a performance gap absent on large-area benchmarks such as Cambridge (SALAD: 0.987) and Tokyo (BoQ: 0.975). The remaining gap arises from sub-20 m class separation and severe inter-class visual ambiguity, confirming that MMS-VPR represents a genuinely harder fine-grained benchmark

RQ2 — Modality Ablation (208 Classes, Field-Collected Data, ResNet Frozen Backbone)

Setup

To isolate the contribution of each modality independently of model capacity, we use a frozen ResNet backbone with a trained linear classifier head, evaluated on all 208 classes of the field-collected dataset (image–video–text–graph aligned). Image and video features are extracted from the frozen ResNet encoder; text features from a frozen pretrained BERT encoder. Fusion weights are optimized jointly with the classifier head.

Split: 80/10/10 per-location stratified; 3 random seeds; early stopping on validation accuracy (patience: 8 epochs). Optimizer: Adam, lr = 1×10⁻³, batch size 64.

Results

Results are means ± std over 3 random splits.

Modality Combination	Accuracy	Macro F1	R@1	R@5	R@10
Image only	0.789 ± 0.001	0.772 ± 0.002	0.789	0.950	0.975
Text only	0.807 ± 0.004	0.762 ± 0.002	0.805	0.988	1.000
Video only	0.947 ± 0.004	0.926 ± 0.007	0.947	0.989	0.996
Image + Text	0.943 ± 0.003	0.923 ± 0.003	0.943	0.999	1.000
Image + Video	0.947 ± 0.004	0.926 ± 0.007	0.947	0.989	0.996
Text + Video	0.981 ± 0.004	0.972 ± 0.005	0.981	1.000	1.000
Image + Text + Video	0.981 ± 0.003	0.971 ± 0.005	0.981	1.000	1.000

Key findings:

• Video is the most powerful single modality: 0.947 vs. 0.789 for image-only (+20.0%) — temporal continuity and motion dynamics provide information unavailable in static frames
• Text excels at coarse retrieval: text alone achieves R@5 = 0.988, meaning OCR-extracted store names and signage are sufficient to place the correct location in the top 5 in ~99% of test cases
• Full multimodal fusion achieves 0.981 — a +19.2% gain over image-only and +3.4% over video-only, as text resolves fine-grained ambiguity between visually similar locations with distinct commercial identities

RQ3 — Cross-Dataset Validation of MMS-VPRlib

MMS-VPRlib reproduces state-of-the-art results on five standard VPR benchmarks, confirming reliable and fair evaluation. Best / second-best per dataset in bold / italic.

Model	Tokyo Acc	New College Acc	Pittsburgh Acc	Nordland Acc	Cambridge Acc
PatchNetVLAD	0.846	0.144	0.870	0.430	0.569
R2Former	0.771	0.722	0.910	0.710	0.946
ViT	0.606	0.641	0.701	0.523	0.815
BLIP	0.674	0.686	0.761	0.583	0.818
MixVPR	0.789	0.758	0.904	0.730	0.971
EigenPlaces	0.927	0.773	0.920	0.681	0.968
ResNet	0.794	0.775	0.879	0.861	0.968
CosPlace	0.816	0.781	0.900	0.560	0.974
CLIP	0.843	0.790	0.901	0.742	0.930
BoQ	0.975	0.787	0.920	0.817	0.984
SALAD	0.804	0.844	0.611	0.744	0.987

Performance inversion: CosPlace achieves 0.974 on Cambridge but only 0.933 on MMS-VPR; SALAD achieves 0.987 on Cambridge but only 0.870 on MMS-VPR. This confirms that fine-grained pedestrian-scale recognition poses distinct challenges from large-area outdoor benchmarks and requires dedicated evaluation — precisely what MMS-VPR is designed for.

---

Use Cases

Visual Place Recognition

• Street-level localization in pedestrian-only and mixed urban environments
• Cross-temporal place matching (day–night, seasonal, multi-year)
• Viewpoint-invariant and illumination-robust recognition

Multimodal Learning

• Vision–language models for place recognition (CLIP, BLIP, and beyond)
• Video-based motion-aware VPR with temporal architectures (LSTM, 3D-CNN, Video Transformers)
• Text-guided visual retrieval using semantic store-name and signage features

Graph Neural Networks

• Spatial relationship modeling in pedestrian networks
• GNN-based place recognition with graph-structured location labels
• Graph-constrained retrieval and spatial adjacency re-ranking

Urban Analytics

• Pedestrian flow analysis using space syntax metrics
• Long-term commercial district characterization from social media
• Spatial accessibility studies in open-air commercial environments

Augmented Reality & Robotics

• Pedestrian AR navigation in dense commercial districts
• Service robot localization in crowd-dense environments
• Context-aware wayfinding at fine spatial granularity

---

Privacy and Ethics

All data collection and release adhere to standard research ethics guidelines.

Privacy Protection

• ✅ Public outdoor spaces only: no private/restricted areas
• ✅ Face detection and pixelation: Automated detection applied to all images and video frames, followed by full manual review to verify complete coverage
• ✅ License plate masking: All vehicle license plates automatically detected and blurred
• ✅ Resolution standardization: All images downsampled to a maximum longest side of 256 pixels (images: 256×192; video frames: 256×144) — at this resolution, individual re-identification is technically infeasible even without face blurring, consistent with Mapillary Vistas practice
• ✅ Layered protection: Face blurring + plate masking + low resolution provide redundant privacy safeguards

Ethical Data Collection

• ✅ Institutional permission: Official permission obtained from Chengdu Taikoo Li management for non-commercial academic data collection and dataset publication
• ✅ Public spaces only: Collection strictly limited to publicly accessible outdoor pedestrian areas; no private, restricted, or indoor spaces
• ✅ Pedestrian-level viewpoints: Natural eye-level capture, not elevated surveillance-style angles
• ✅ Normal operating hours: Data collection during 7AM–10PM only; no late-night intrusion
• ✅ Systematic and unbiased: Spatial and temporal coverage designed to avoid sampling bias
• ✅ Manual review: All processed media manually reviewed post-anonymisation to verify complete PII removal

Dual-Use Mitigation

MMS-VPR is designed to advance research in navigation, augmented reality, and robotics — not surveillance. Safeguards include CC BY-NC-SA 4.0 licensing prohibiting commercial use, resolution capping making re-identification technically infeasible, and explicit guidance in this data card discouraging downstream surveillance applications.

Note: While automated anonymisation and manual review have been applied, if you identify any remaining privacy concerns, please contact us immediately via GitHub Issues or email.

Social Media Data

All Weibo-sourced images are: (1) publicly posted content on the Weibo platform; (2) fully anonymized via the pipeline above; (3) restricted to non-commercial academic use under CC BY-NC-SA 4.0. As a fallback for practitioners requiring unrestricted redistribution, we provide precomputed CLIP ViT-B/32 embeddings (512-dim) for the Weibo subset, preserving multimodal utility without redistribution concerns. The field-collected subset alone constitutes a self-contained, fully functional multimodal VPR benchmark.

---

License

This dataset is released under the Creative Commons Attribution–NonCommercial–ShareAlike 4.0 International (CC BY-NC-SA 4.0) license. This applies uniformly to both field-collected and Weibo-sourced imagery.

You are free to:

• Share: Copy and redistribute the material in any medium or format
• Adapt: Remix, transform, and build upon the material

Under the following terms:

• Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made
• NonCommercial: You may not use the material for commercial purposes
• ShareAlike: If you remix or build upon the material, you must distribute your contributions under the same CC BY-NC-SA 4.0 license

Field-collected data (78,575 images, 2,527 videos): collected with permission from Chengdu Taikoo Li management for non-commercial academic research.

Weibo-sourced data (31,954 images): publicly posted content on Weibo, anonymised before release, and restricted to non-commercial academic use consistent with the CC BY-NC-SA 4.0 license.

As a redistribution-safe alternative, precomputed CLIP ViT-B/32 embeddings (512-dim) for the Weibo subset are available upon request. The field-collected subset alone constitutes a fully functional multimodal VPR benchmark.

---

Acknowledgments

We thank all contributors to this dataset. Field data were collected by the research team in 2024 with official permission from Chengdu Taikoo Li. Social media data were curated from publicly shared posts on Weibo (2019–2025).

---

Citation

If you use MMS-VPR or MMS-VPRlib in your research, please cite:

@article{ou2025mmsvpr,
  title     = {MMS-VPR: A Fine-Grained Multimodal Street-Level Visual Place Recognition Dataset and Evaluation Benchmark for Dense Pedestrian Environments},
  author    = {Ou, Yiwei and Ren, Xiaobin and Sun, Ronggui and Gao, Guansong and Zhao, Kaiqi and Manfredini, Manfredo},
  journal   = {arXiv preprint arXiv:2505.12254},
  year      = {2025},
  url       = {https://arxiv.org/abs/2505.12254}
}

---

Contact

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

• GitHub Issues: MMS-VPRlib Repository
• Email: you661@aucklanduni.ac.nz
• Hugging Face: Dataset page

---

Updates

Latest Version: v3.0 (May 2026)

• Benchmark expanded to 22 baseline models (from 17)
• Added cross-dataset validation on 5 external VPR benchmarks (Tokyo, Pittsburgh, Nordland, Cambridge, New College)
• Added full modality ablation results (image / text / video / graph combinations; full table in paper)
• License updated to CC BY-NC-SA 4.0 (field and social media data unified under same license)
• Added Sample Dataset folder for rapid data quality inspection without full download
• Expanded privacy and ethics documentation; added written-permission statement
• Updated paper title and author list to match final version

---

v2.0 (February 2026)

• 208 locations (updated from 207)
• Integrated social media data (2019–2025)
• Enhanced textual annotations
• Complete graph structure documentation