Accuracy Issues

Accuracy & Results Issues

Understand confidence scores, interpret prediction quality, and learn when to trust results. This guide helps you assess accuracy and improve prediction reliability.

Updated 2 weeks ago

01
Confidence

Understanding Confidence Scores

Every prediction includes a confidence percentage indicating how certain the AI is about the location. Use these scores to evaluate result quality.

85+

High Confidence

Excellent reliability. Results are typically accurate within a few blocks.

60-84

Moderate Confidence

Reasonable accuracy, verify using map view. Results are typically accurate to neighborhood level.

<60

Low Confidence

Uncertain prediction. Treat these results as rough estimates only.

What Affects Confidence

  • -Visible text: Street signs, business names, and landmarks increase confidence
  • -Architectural uniqueness: Distinctive buildings improve accuracy
  • -Image clarity: Sharp, well-lit photos produce better results
  • -Street-level view: Ground perspective provides more geographic clues
02
Troubleshooting

Low Accuracy Predictions

When confidence scores are low, the prediction may be unreliable. Understand why this happens and what to do about it.

Generic Architecture

Buildings and streets that look similar across multiple cities confuse the model. Modern glass towers, residential suburbs, and generic storefronts lack distinctive features.

Solutions

  • -Look for unique landmarks or signage in different parts of the image
  • -Try analyzing a different angle of the same location
  • -Test the image with multiple city models to compare results

Insufficient Visual Information

Close-ups, cropped views, or images focused on a single object don't provide enough context for accurate geolocation.

Solutions

  • -Use wider-angle photos that show street context and surrounding area
  • -Include multiple buildings or landmarks in frame
  • -Ensure street signs or business names are visible

Wrong City Model Selected

Using the Miami model for a New York photo produces low-confidence results since each model is trained on specific city data.

Solutions

  • -Try all four city models if you're uncertain about the location
  • -The model with the highest confidence score is usually correct
  • -Look for city-specific clues (architecture style, vegetation, street layouts)
03
Verification

Verification Strategies

Always verify predictions before relying on them, especially for moderate to low confidence scores.

Use Map View Comparison

After receiving a prediction, compare the map view with your original image. Look for matching:

  • -Building shapes and arrangements
  • -Street patterns and intersections
  • -Landmarks or distinctive structures
  • -Waterfront, parks, or geographic features

Check Street View

Use the street view feature in mapping tools to visually confirm the location. This is the most reliable verification method-if the street view matches your image, the prediction is accurate.

Cross-Reference Text Elements

If your image contains business names, street signs, or other readable text, search for them online combined with the predicted city. This confirms whether those businesses or streets exist in that location.
04
Limitations

When Predictions Are Unreliable

Certain scenarios produce inherently unreliable results, regardless of model quality.

Do Not Trust Results For:

  • -Indoor photos: Office interiors, restaurant dining rooms, or building lobbies lack external geographic context
  • -Aerial/drone shots: High-altitude views don't capture street-level details the model is trained on
  • -Close-up objects: Photos of individual items, food, people, or products
  • -Heavy filters/editing: Heavily processed images that alter colors, add effects, or obscure details
  • -Historical photos: Images from decades ago when buildings and streets looked different
  • -Natural landscapes: Parks, beaches, mountains without visible man-made structures

Important

OceanIR is designed for street-level urban photography showing buildings, roads, and city infrastructure. Using it for other image types produces random or meaningless results.
05
Best Practices

Improving Result Quality

Get better predictions by following these best practices for image selection and analysis.

Ideal Images

  • -Street view from ground level
  • -Multiple buildings visible
  • -Street signs or business names
  • -Clear, well-lit, unedited photos
  • -Distinctive architecture

Avoid

  • -Blurry or low-resolution images
  • -Heavy Instagram filters
  • -Extreme close-ups or crops
  • -Night photos with poor lighting
  • -Generic suburban areas

Pro Tip

If you have multiple photos of the same location, analyze the one with the clearest view of street signs and landmarks. Even a slightly different angle can significantly improve accuracy by capturing additional context.