Amazon Textract adapter for Custom Query seems to not use it's training.

Question

I have an adapter that I trained on a query that looks for a logo on every page. It returns a binary response. When I trained the adapter, I got 100.0% accuracy on all pages. I had a training/test split of 93/77 pages. Now, I am attempting to use the adapter on some documents that were in the test set and some documents that were not. About half the responses are wrong. Has anyone else run into this issue, and if so, how did you solve it?

Answer

Hi Jon,

Thank you for your question. 
* If you're looking to detect brand logos that contain a combination of text and logos, we recommend trying Amazon Rekognition Custom Labels.
  * https://aws.amazon.com/blogs/machine-learning/part-1-end-to-end-solution-building-your-own-brand-detection-and-visibility-using-amazon-sagemaker-ground-truth-and-amazon-rekognition-custom-labels/
* Amazon Textract and Custom Labels is a good fit for OCR extraction, for e.g. if your use case is to extract or detect presence of underlying text, in which case, we recommend adding more training samples that are representative of your overall sample set.
  * https://docs.aws.amazon.com/textract/latest/dg/best-practices-adapters.html

Hope this answers your question, and if so, please accept this answer.
Have a great day!

Amazon Textract adapter for Custom Query seems to not use it's training.

Relevant content