HERE Hackathon: POI Validation on Multiply-Digitised Roads

🚀 Inspiration

Learning new technologies and solving challenges that motivate us to keep on growing as developers and growing our network by working as a team and sharing knowledge with each other.

🤖 What it does

This project automatically analyzes POIs located on multiply-digitised roads to determine if they:

Actually exist in reality.
Are on the correct side of the road.
Are legitimate exceptions (e.g., truly between lanes).
Or are errors due to misclassification or data issues.

For each POI, the system:

Interpolates its exact location along the road geometry.
Generates and rotates satellite image patches centered on the POI.
Uses a CLIP-based AI model to classify the patch and determine if a POI is visible, and on which side.
Assigns a scenario and suggests an action (e.g., delete, move, or keep the POI).
Outputs results and suggested actions to a CSV for easy review.

🏗️ How we built it

Data Processing: Loads and indexes HERE's NAV and Naming GeoJSONs and POI CSVs.
Geometric Analysis: Filters POIs on MULTIDIGIT = 'Y' roads, interpolates their position, and builds polygons between forward/backward lanes.
Satellite Imagery: Downloads and rotates satellite tiles from HERE Maps, generating image patches for each POI.
AI Classification: Uses OpenAI's CLIP model to classify each patch as containing a POI or not, and to determine its position (center or sidewalk).
Scenario Assignment: Applies business rules to assign each POI to one of the following scenarios:
1. Non-existent POI → Suggest Delete POI
2. Wrong side of the road → Suggest Change Side
3. Legitimate exception → No action needed
Output: Saves results and actions to a CSV file and generates example images for flagged POIs.

🧩 Challenges we ran into

Satellite Alignment: Ensuring patches are correctly rotated and centered on the POI, regardless of road orientation.
Ambiguous POIs: Some POIs are visually ambiguous or not visible in satellite imagery, making AI classification challenging.
Data Quality: Handling inconsistencies in the source data, such as missing or misclassified road segments.
Performance: Processing large datasets efficiently, especially when downloading and handling high-resolution imagery.

🏅 Accomplishments that we're proud of

Fully automated pipeline from raw data to actionable CSV output.
Robust geometric analysis to handle complex road layouts.
Integration of zero-shot AI classification (CLIP) for visual POI validation.
Clear scenario tagging and action suggestion for each POI.
Modular codebase that can be extended to other cities or datasets.

📚 What we learned

The importance of combining geometric, visual, and semantic analysis for robust map validation.
How to leverage zero-shot models like CLIP for geospatial tasks.
The value of clear scenario definitions and business rules in automating map data QA.

🔮 What's next for POI Validation

Model Improvements: Fine-tune or ensemble models for even better POI detection.
UI/Visualization: Build a web dashboard for interactive review and correction of flagged POIs.
Scalability: Optimize for batch processing of entire cities or countries.
Generalization: Extend to other types of map features (e.g., traffic signs, crosswalks).
Feedback Loop: Integrate user feedback to continually improve classification accuracy.