MosaicFusion: Diffusion Models as Data Augmenters for Large Vocabulary Instance Segmentation

Jiahao Xie¹  Wei Li¹  Xiangtai Li¹  Ziwei Liu¹  Yew Soon Ong²  Chen Change Loy¹

¹S-Lab, ²Nanyang Technological University

🚩 Accepted to IJCV 2024

• [arXiv] •

We present MosaicFusion, a general diffusion-based data augmentation pipeline for large-vocabulary instance segmentation. The MosaicFusion-synthesized instance segmentation dataset can be used to train various downstream detection and segmentation models to improve their performances, especially for rare and novel categories.

🤩 Key Properties

Training-free Directly generate multiple objects Agnostic to detection architectures Without extra detectors or segmentors

---

MosaicFusion is a training-free diffusion-based dataset augmentation pipeline that can produce image and mask pairs with multiple objects simultaneously using the off-the-shelf text-to-image diffusion models. The overall pipeline of MosaicFusion consists of two components: image generation and mask generation.

Given only interest category names, MosaicFusion can generate high-quality multi-object images and masks simultaneously by conditioning on a specific text prompt for each region.

• Clone our repo from GitHub:

git clone https://github.com/Jiahao000/MosaicFusion.git
cd MosaicFusion

• Create the conda environment:

conda env create -f environment.yml

• Download lvis_v1_train.json, unzip and put it under a directory, e.g., data/lvis/meta/lvis_v1_train.json.

1. Generate images and masks with MosaicFusion:

bash scripts/dist_text2seg.sh "a photo of a single category" output/text2seg Generation_log

Alternatively, if you run MosaicFusion on a cluster managed with slurm:

bash scripts/slurm_text2seg.sh Dummy Generation_job "a photo of a single category" output/text2seg Generation_log

2. Convert generated images and masks to the required data format:

bash scripts/run_seg2ann.sh output/text2seg output/seg2ann

3. Merge MosaicFusion annotations into LVIS annotations:

bash scripts/run_merge_ann.sh data/lvis/meta/lvis_v1_train.json output/seg2ann/annotations/lvis_v1_train_mosaicfusion.json output/seg2ann/annotations/lvis_v1_train_merged.json

Please refer to TRAIN.md for training details.

• Data generation code for MosaicFusion
• Third-party training code with MosaicFusion data

If you find this work useful for your research, please consider citing our paper:

@article{xie2024mosaicfusion,
  author = {Xie, Jiahao and Li, Wei and Li, Xiangtai and Liu, Ziwei and Ong, Yew Soon and Loy, Chen Change},
  title = {MosaicFusion: Diffusion Models as Data Augmenters for Large Vocabulary Instance Segmentation},
  journal = {International Journal of Computer Vision},
  year = {2024}
}

Distributed under the S-Lab License. See LICENSE for more information.