The Image Evaluation Threshold is a critical parameter used in search specifications for product-type searches (e.g., chairs, lamps, toothbrushes). It controls how closely a finding's image must match your reference images.
How the Image Evaluation Threshold Works
When a Marketplace Search Specification is configured with reference images of your product, the system uses an image recognition algorithm to compare all findings’ images against those reference images. Each image comparison results in a similarity score, ranging from 0 to 100, where:
- 100 represents a perfect 1:1 match between the images.
- 70-100 represents near-identical matches.
- 50-60 represents broad matches, where there are clear similarities but with slight variations.
- 30-40 represents very broad matches, where the images share only general features.
The Image Evaluation Threshold sets the minimum score required for a finding to proceed to the platform. If the similarity score is above the threshold, the finding is uploaded. If it’s below the threshold, the finding is ignored.
Default Threshold and Adjustments
The default threshold is set at 50, which captures broad matches—ideal for early iterations of a search. Here’s how you can adjust the threshold to fine-tune the results:
- 30-40: These thresholds capture very broad matches, making it useful for initial searches where you want to cast a wide net and catch all variations of your product images.
- 50: This is the default and captures broad but relevant matches. It’s a balance between inclusivity and specificity.
- 60-70: For searches that have already gathered several verified findings, these thresholds capture very close matches, reducing noise in the results.
- 70+: For highly specific and nearly identical matches. Only images that are almost exactly like your reference images will be flagged.
Why Adjusting the Threshold is Important
Fine-tuning the Image Evaluation Threshold allows you to control the quality of the findings:
- Lower thresholds (30-50) are useful when you first run a search. They help capture a wide variety of potential infringements, even if the images are not exact matches.
- Higher thresholds (60-70) can be used once the system has learned from previous, verified results. This helps reduce irrelevant findings (noise) and focuses on precise matches, making enforcement faster and more efficient.
Practical Use Case
Let’s say you’re searching for counterfeit versions of a designer chair, and you’ve included images of the authentic chair in your search specification. After a search runs:
- A finding from Amazon.de includes an image that scores 75 compared to your reference image. Since this is above the default threshold of 50, the finding proceeds to be uploaded for review and potential enforcement.
- A finding on Taobao scores 35. This falls below the threshold, so it is discarded, preventing irrelevant results from being added to the platform.
By adjusting the threshold over time, you can ensure that the system captures relevant matches while filtering out irrelevant or too-broad results.
Tips for Adjusting the Threshold
- Initial Searches:
Start with a lower threshold (30-40) for early searches to catch as many image variations as possible.
- After Verification:
Once you’ve verified several findings, gradually increase the threshold to focus on more accurate, high-quality results (60-70).
- Long-Term Optimization:
Regularly revisit and adjust the threshold as the system learns and new findings are generated. This allows you to maintain the balance between capturing potential infringements and filtering out unnecessary results.
Conclusion
The Image Evaluation Threshold is a powerful tool that allows you to customize the sensitivity of your image searches. By adjusting the threshold based on the stage of your search and the precision you need, you can optimize the balance between capturing broad potential matches and reducing irrelevant findings.
For more details on configuring your Image Evaluation Threshold or optimizing search specifications, please see our article on Marketplace Search Specifications.