Generative-AI data augmentation:
Objective:
The Visual Query Tool solved the problem of manual data curation by automating it. But another problem arises when you dont have data in your database similar to your query. To address this, I developed a Diffusion model based data augmentation pipeline to generate images similar to the query.
Contributions:
- Trained an image2image diffusion model using Adobe affordance insertion method ((Affordance Insertion)) for image prompt based inpainting.
- Generated new images for training YoloNAS model on failure cases with less representation in the dataset.
Result:
- Developed a method for generating new images using image propmts.
- Enhanced the YoloNAS model's F1-score for obstacle detection by 18% through fine-tuning with newly generated data.