Introduction

Datagen(Data synthesis) allows the synthesis of images based on a few sample images. This is especially useful for rare data types (e.g., defects on a metal surface) that are challenging to collect. The following diagram shows the overall datagen flow in Data Explorer.

1. Provide inputs that consist of

   1. Concept of interest images: Draw bounding boxes to mark the concept to be synthesized. At least six images are required.

   2. Generation canvas: Draw polygons to mark the location on the image where the concept must be synthesized. At least six images are required.

   3. Background images: Provide image samples representing the variety of backgrounds over which the concept must be synthesized. Fifty background images are recommended.

2. Run a preview session.

3. Review the results

   1. If the results are satisfactory, then ‘Accept’ the underlying specification.

   2. If the results are not satisfactory, click ‘Refine’ to open a new version of the specification. Provide feedback as shown below.
      

4. Run the preview session and repeat the review process.

5. Once a specification is ‘Accept’ed, ‘Generate’ to initiate one or more generation sessions.