Terangavoice/Teranga_Ground_Truth

Dataset Card for Teranga Ground Truth (TGT) v1.0 - Evaluation Sample

Dataset Details

Dataset Description

Teranga Ground Truth (TGT) v1.0 is a high-resolution, georeferenced urban imagery and Point of Interest (POI) dataset collected across key municipal districts in Dakar, Senegal. Engineered specifically to solve the critical "Domain Shift" problem in global Computer Vision models deployed in West Africa, TGT captures ground-level visual realities including informal commerce, unique road infrastructure, local building density, and atmospheric light conditions.

This repository hosts the 500-photo Evaluation Sample designed for academic research, algorithm testing, and data validation. The complete commercial dataset (7,500 high-resolution images) includes enriched tabular CSV metadata, standardized GeoJSON layers (RFC 7946), an interactive HTML map, and enterprise-grade legal compliance documentation.

• Curated by: DLO GLOBAL (Dakar, Senegal)
• Funded by: DLO GLOBAL
• Shared by: DLO GLOBAL Data Engineering Team
• Language(s): English (Metadata/Documentation), French/Wolof (On-street signage/text)
• License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)

Dataset Sources

• Repository: Kaggle Evaluation Pack
• Interactive Demo & Licensing: Contact DLO GLOBAL for full enterprise access.

---

Uses

Direct Use

This evaluation pack can be directly used for:

• Evaluating Computer Vision model robustness against West African urban visual features (Domain Shift analysis).
• Ground truth verification and calibration for satellite and remote sensing data.
• Academic research on computer vision fairness, urban morphology, and geographical AI equity.
• Testing spatial indexing and GeoJSON pipeline integrations.

Out-of-Scope Use

• Commercial Production Deployment: Using this evaluation pack or derived models in commercial products without purchasing a DLO GLOBAL Commercial License is strictly prohibited under CC BY-NC 4.0.
• Surveillance or Re-identification: Attempting to de-anonymize blurred faces, license plates, or personal identifiers.

---

Dataset Structure

The evaluation dataset contains high-resolution images paired with a structured index file:

• Images: High-resolution JPEG files captured at ground level across 5 municipal zones in Dakar.
• Tabular Metadata (metadata.csv):
  • image_id: Unique identifier for each frame.
  • latitude / longitude: Exact WGS84 coordinates.
  • timestamp_utc: Standardized UTC acquisition timestamp.
  • camera_parameters: Optical metadata and sensor orientation information.
• Spatial Vector File (layers.geojson): RFC 7946 compliant GeoJSON features for seamless GIS integration (QGIS, ArcGIS, Mapbox).

---

Dataset Creation

Curation Rationale

Global AI models frequently fail in African urban contexts due to the lack of localized, high-quality ground-level datasets. TGT v1.0 was created to establish a benchmark data infrastructure for West Africa, empowering researchers and enterprises to build unbiased, spatially accurate models.

Source Data

Data Collection and Processing

Data was systematically collected across 5 strategic municipal districts in Dakar, Senegal, using calibrated optical equipment mounted at ground level. Capture routes were mapped to ensure balanced spatial distribution covering main avenues, commercial corridors, and residential neighborhoods.

Who are the source data producers?

The data was collected, curated, and processed directly by the data engineering team at DLO GLOBAL, headquartered in Dakar, Senegal.

Annotations

Annotation process

Anonymization processing is executed via a two-stage hybrid pipeline:

1. Automated AI blurring passes for human faces and vehicle license plates.
2. 100% human Quality Assurance (QA) verification to ensure strict compliance with local and international privacy regulations.

Personal and Sensitive Information

The dataset strictly complies with Senegalese Personal Data Protection Regulations (Loi CDP n° 2008-12) and aligns with global GDPR standards for public space data collection. All identifiable personal data (faces and vehicle registration plates) have been blurred.

---

Bias, Risks, and Limitations

Limitations

• Geographic Scope (v1.0): Data collection for v1.0 is limited to 5 municipal zones within the Dakar peninsula and may not fully represent rural or secondary urban centers in West Africa.
• Temporal Scope: Captures represent specific seasonal and lighting conditions corresponding to the v1.0 acquisition timeframe.

Recommendations

Users should account for geographical specificity when evaluating model generalization. To train production-grade models or expand coverage, enterprise users should inquire about full dataset licensing and custom collection pipelines.

---

Citation

If you use this dataset in your research or benchmark evaluations, please cite it as follows:

@dataset{dlo_global_tgt_2026,
  author       = {DLO GLOBAL},
  title        = {Teranga Ground Truth (TGT) v1.0: Benchmark Ground-Level Urban Imagery Dataset for West Africa},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{[https://huggingface.co/datasets/DLOGLOBAL/teranga-ground-truth-v1](https://huggingface.co/datasets/DLOGLOBAL/teranga-ground-truth-v1)}},
  note         = {Evaluation Sample under CC BY-NC 4.0}
}