Awesome

Introduction
Requirements
Usage
- Generate synthetic data
- Train the semantic segmenter
Contacts

PyTorch implementation of Dataset Diffusion: Diffusion-based Synthetic Data Generation for Pixel-Level Semantic Segmentation (NeurIPS2023)

Quang Nguyen, Truong Vu, Anh Tran, Khoi Nguyen<br> VinAI Research, Vietnam

Abstract: Preparing training data for deep vision models is a labor-intensive task. To ad- dress this, generative models have emerged as an effective solution for generating synthetic data. While current generative models produce image-level category labels, we propose a novel method for generating pixel-level semantic segmen- tation labels using the text-to-image generative model Stable Diffusion (SD). By utilizing the text prompts, cross-attention, and self-attention of SD, we introduce three new techniques: class-prompt appending, class-prompt cross-attention, and self-attention exponentiation. These techniques enable us to generate segmentation maps corresponding to synthetic images. These maps serve as pseudo-labels for training semantic segmenters, eliminating the need for labor-intensive pixel-wise annotation. To account for the imperfections in our pseudo-labels, we incorporate uncertainty regions into the segmentation, allowing us to disregard loss from those regions. We conduct evaluations on two datasets, PASCAL VOC and MSCOCO, and our approach significantly outperforms concurrent work.

Details of the model architecture and experimental results can be found in our following paper.<br> Please CITE our paper whenever this repository is used to help produce published results or incorporated into other software.

@inproceedings{quangtruong2023@dd,
  title={Dataset Diffusion: Diffusion-based Synthetic Dataset Generation for Pixel-Level Semantic Segmentation},
  author={Quang Ho Nguyen, Truong Vu, Anh Tran, Khoi Nguyen},
  year={2023},
  booktitle={Thirty-Seventh Conference on Neural Information Processing Systems},
}

Requirements

OS: we recommend Linux for performance and compatibility reasons.
64-bit Python ≥ 3.8
PyTorch ≥ 1.10 and torchvision that matches the PYTorch installation. Follow official instruction
HuggingFace installations: diffusers, transformers, safetensors
pip install --user -U nltk

Please refer to INSTALL.md for installations of MMCV and datasets preparation.

Usage: Command lines

Generating synthetic data

To generate synthetic dataset with VOC classes, run

sh scripts/gen_data_voc.sh

The dataset will be saved to data/gen_voc by default.

To generate synthetic dataset with COCO classes, run

sh scripts/gen_data_coco.sh

The dataset will be saved to data/gen_coco by default.

Training the semantic segmenter

To train the semantic segmenter on VOC synthetic dataset, run

sh scripts/train_mmseg_voc.sh

To train the semantic segmenter on COCO synthetic dataset, run

sh scripts/train_mmseg_coco.sh

You can alter the number of GPUs for training specified in these scripts.

Dataset

The generated dataset is provided here

By downloading this dataset, USER agrees:

to use the dataset for research or educational purposes only.

to not distribute the dataset or part of the dataset in any original or modified form.

and to cite our paper whenever the dataset is used to help produce published results.

Contacts:

If you have any questions about this project, please don't hestitate to drop me an email quangngcs@gmail.com or open an issue in this repository