Introduction to the project

Image segmentation with AI techniques. - MSc. Saúl Cano-Ortiz (University of Cantabria)

1. Statement of the problem: 2D segmentation challenges.
2. State-of-the-art segmentation architectures.
3. Possible solution: Data augmentation with synthetic simples.
4. State-of-the-art generative models applied to semantic image synthesis.
5. A real-world application: intelligent road maintenance.
6. Main conclusion.
   
   

Statement of the problem: 2D segmentation challenges

State-of-the-art segmentation architectures

Data augmentation with generative architecture.

You have to label really few images and then once the generative model is trained, if it works, you can generate thousands/as many images as you want.

Generative models for semantic segmentation.

You can create thousands of images in seconds or milliseconds.

CycleGAN

cycle-consistency method - instead of training two networks, you have two networks (two generators and two discriminators which is more complex)

Diffusion Models
denoising of the images to get the final result

Instead of giving us input noise what you get is the output of an encoder. You encode the noise so the dimension is much smaller.

A real-world application: intelligent road maintenance

Main conclusion

• Segmentation architectures are data-driven.
• It is interesting to focus on Neural Architecture Search.
• But remember, it is also about data quality and label creation.
• Create easily synthetic frames from hand-painted pictures.
• Augment the dataset, and then explore, segmentation architectures.