Accuracy Disclaimer

1. Overview

GeoPin provides geolocation inference for the Netherlands using machine learning models. This disclaimer explains the inherent limitations of our service and the appropriate use of geolocation results.

By using the GeoPin service, you acknowledge that you have read and understood this disclaimer, and you agree that GeoPin cannot be held liable for decisions, actions or consequences arising from reliance on geolocation results without independent verification.

2. Nature of Results

GeoPin's geolocation results are probabilistic estimates, not verified facts. Each result represents the model's best prediction of where an image was taken, based on visual features and patterns learned during training. Key characteristics:

• Confidence scores are relative, not absolute. A confidence score of 85% does not mean there is an 85% probability that the predicted location is correct. Confidence scores indicate the model's relative certainty compared to other possible locations and should be interpreted as ordinal rankings, not calibrated probabilities.
• Results are estimates with a margin of error. Even high-confidence predictions may be off by tens of metres, hundreds of metres or more, depending on the specificity of the visual features in the image.
• Results can be entirely incorrect. The model can produce a confidently incorrect prediction. This is an inherent property of machine learning systems and does not indicate a defect in the service.
• Results do not provide certainty. A geolocation result should always be treated as a hypothesis to be verified, not as a conclusion to be relied upon.

3. Model Limitations

The machine learning models used by GeoPin have inherent limitations that affect the accuracy and reliability of results:

3.1 Training Data Limitations

• Models are trained on a finite dataset of geo-tagged images from the Netherlands. The dataset may not represent all regions, seasons, lighting conditions or types of environments uniformly.
• Areas with fewer training images (rural areas, newly developed areas, recently renovated buildings) may yield less accurate results.
• The training data has a temporal cutoff. Changes to the physical environment that have occurred after the training data was collected (new construction, demolished buildings, changed signage) are not reflected in the model.

3.2 Visual Ambiguity

• Many locations in the Netherlands share similar visual characteristics (e.g. canal houses, polder landscapes, standardised road infrastructure). The model may confuse visually similar but geographically different locations.
• Generic or feature-sparse images (e.g. interior photos, close-ups, uniformly flat landscapes without distinguishing features) provide insufficient information for accurate geolocation.

3.3 Adversarial and Edge Cases

• Images that have been heavily edited, filtered, distorted or artificially generated may produce unpredictable results.
• The model is designed for photographs of real locations. Results for paintings, drawings, screenshots, satellite imagery or synthetic images are not reliable.
• Images deliberately designed to mislead the model (adversarial examples) may produce incorrect results with high confidence scores.

3.4 Geographic Scope

• GeoPin is optimised for the Netherlands. Images depicting locations outside the Netherlands may produce incorrect results pointing to Dutch locations, or may return no result.
• Border areas may produce results on either side of the actual border with Belgium or Germany.

4. Factors Affecting Accuracy

The accuracy of GeoPin results is influenced by various factors, many of which are beyond our control:

5. Not Legal Evidence

GeoPin results are expressly not intended as legal evidence. You may not:

• Present GeoPin results as evidence in court proceedings, arbitration or other legal forums without independent verification by a qualified expert.
• Use GeoPin results as the sole basis for filing criminal complaints, civil claims or administrative proceedings.
• Use GeoPin results to establish or dispute a person's presence or absence at a particular location for legal purposes.
• Cite GeoPin results in sworn statements, deeds or notarial documents without appropriate qualification and independent corroboration.

If you intend to use geolocation data in a legal context, you should engage a qualified forensic analyst or expert witness who can independently verify the location through multiple methods and testify about the methodology and its limitations.

6. Data Sources & Attribution

GeoPin's machine learning models have been developed using publicly available datasets of geo-tagged images, including but not limited to:

• Publicly available street-level imagery and geo-tagged photographs.
• Open geodata published by Dutch government agencies and municipalities.
• Open-source geographic datasets and map data.

GeoPin does not claim ownership of the underlying geographic data or images used for model training. GeoPin's geolocation results are derived from the model's learned representations and are not direct reproductions of source material.

If you are a rights holder and believe that GeoPin's training data contains material that infringes your rights, please contact us at info@geopin.nl.

7. No Warranty

To the maximum extent permitted by Dutch law:

• GeoPin provides its service and geolocation results "as is" and "as available" without any warranty, whether express, implied or statutory.
• GeoPin specifically disclaims any warranty of accuracy, completeness, reliability, fitness for a particular purpose or non-infringement with respect to geolocation results.
• GeoPin does not warrant that results will be error-free, uninterrupted or suitable for your specific requirements.
• No oral or written information or advice given by GeoPin or its representatives creates a warranty not expressly stated in our Algemene Voorwaarden.

You accept all risks associated with your use of GeoPin results. GeoPin is not liable for any loss, damage or injury arising from reliance on geolocation results, including but not limited to financial loss, reputational damage or personal injury.

8. Recommended Practices

To get the most reliable results from GeoPin and use them responsibly, we recommend the following practices:

8.1 Verification

• Always verify GeoPin results through independent sources before acting on them. Compare with map services, street-level imagery, local knowledge or other geolocation tools.
• Pay attention to the confidence score. Lower confidence scores indicate greater uncertainty and a higher likelihood of errors.
• Consider submitting multiple images of the same location (from different angles, if available) and comparing the results.

8.2 Context

• Provide contextual information where available. If you know the general region (province, city), this can help you assess whether a result is plausible.
• Consider the age of the photo. Older images may show features that no longer exist.

8.3 Reporting and Communication

• When sharing or publishing GeoPin results, always include a disclaimer that the result is an automated estimate and has not been independently verified (unless it has been).
• Include the confidence score and relevant caveats when communicating results to others.
• Do not present GeoPin results with more certainty than is warranted by the confidence score and the limitations described in this document.

9. Contact

If you have questions about this Accuracy Disclaimer or the limitations of the GeoPin service, please contact us: