license: apache-2.0
tags:
- art
pretty_name: Human Segmentation Dataset
Human Segmentation Dataset
This dataset was created for developing the best fully open-source background remover of images with humans. It was crafted with LayerDiffuse, a Stable Diffusion extension for generating transparent images. After creating segmented humans, IC-Light was used for embedding them into realistic scenarios.
The dataset covers a diverse set of segmented humans: various skin tones, clothes, hair styles etc. Since Stable Diffusion is not perfect, the dataset contains images with flaws. Still the dataset is good enough for training background remover models. I created more than 7.000 images with people and diverse backgrounds.
Example
Support
If you identify weaknesses in the data, please contact me.
I had some trouble with the Hugging Face file upload. This is why you can find the data here: Google Drive.
Research
Synthetic datasets have limitations for achieving great segmentation results. This is because artificial lighting, occlusion, scale or backgrounds create a gap between synthetic and real images. A "model trained solely on synthetic data generated with naïve domain randomization struggles to generalize on the real domain", see PEOPLESANSPEOPLE: A Synthetic Data Generator for Human-Centric Computer Vision (2022). However, hybrid training approaches seem to be promising and can even improve segmentation results.
Currently I am doing research how to close this gap. Latest research is about creating segmented humans with LayerDiffuse and then apply IC-Light for creating realistic light effects and shadows.
Changelog
08.06.2024
- Applied IC-Light to segmented data
- Added higher rotation angle to augmentation transformation
28.05.2024
- Reduced blur, because it leads to blurred edges in results
26.05.2024
- Added more diverse backgrounds (natural landscapes, streets, houses)
- Added more close-up images
- Added shadow augmentation