Our client, Click-Ins, works with a dataset of images of vehicles with different types of damages. These vehicles are of varying vehicle models and colors. Damages are of varying types, size, pattern, location. As the images of damaged vehicles are rare and hard to acquire, Click-ins put great emphasis on generating these synthetically. The tricky part, however, has always been the realistic looks of the damaged part, with its irregular shape. This is where our team and generative AI came to help.
Tightly cooperating with the Clickins R&D, we built a solution that generates a sufficient image database, by automatic synthesis of vehicle images with damage structures. We created a tool for synthesizing damages with annotations that can be used in training AI systems. Our Objectives in prototype phase were stability – preserves non-defect texture and application scope – applicable to more car colours & views and defect locations. Primary focus (output) was on defect textures with their silhouettes.
Training dataset of real photos of damaged cars is used during the training of generative models for synthesis of the realistically looking damage structure on the surface of real cars. Filtering was done on Cognexa side according to the annotations provided from the client.
The solution improved the client’s core AI systems, as it provides the supply of high quality annotated synthetic data with unmatched variability.