Ariadna Kramkovska - Background removal without background knowledge

Ariadna Kramkovska presents a proof of concept for background removal without background knowledge, showcasing successes and challenges with U2Net, alpha matting, and post-processing techniques.

Key takeaways

Computer vision falls under the umbrella of salient object detection, which focuses on detecting the human eye’s focus, or the background.
External tools are a good starting point, as they can be easily adopted and provide comfort with use.
The model used was U2Net, a pre-trained neural network with an encoder-decoder structure.
The original image and a black and white mask are used as the ground truth for training.
A dataset of 25 image pairs was created manually for training, with an emphasis on removing backgrounds.
Results showed that U2Net has potential, with 80% accuracy, but it’s not perfect and may require further fine-tuning.
Diffusion models can be used to generate data or masks, but this was not implemented due to computational expense.
The company, Printify, uses background removal for creating customized products, such as t-shirts and mugs.
The approach was to create a cleaner dataset, remove poor-quality images, and use alpha matting to correct for errors.
The project is a proof of concept, with a successful outcome.
The next step is to compare the in-house model with external tools, considering factors such as inference speed and pricing.
Historical data and stable diffusion data can be used in combination to balance out the approach.
The model was trained on a local PC, with a small dataset and quick results, which was surprising.
Post-processing is important, as it can remove details that are not desired.
The company is also using AI images on its website, allowing merchants to create their own designs with AI.

Ariadna Kramkovska - Background removal without background knowledge

More talks