OneFormer: One Transformer to Rule Universal Image Segmentation

Jitesh Jain, Jiachen Li^†, MangTik Chiu^†, Ali Hassani, Nikita Orlov, Humphrey Shi

^† Equal Contribution

[Project Page] [arXiv] [pdf] [BibTeX]

This repo contains the code for our paper OneFormer: One Transformer to Rule Universal Image Segmentation.

<img src="images/teaser.png" width="100%"/>

Features

• OneFormer is the first multi-task universal image segmentation framework based on transformers.
• OneFormer needs to be trained only once with a single universal architecture, a single model, and on a single dataset , to outperform existing frameworks across semantic, instance, and panoptic segmentation tasks.
• OneFormer uses a task-conditioned joint training strategy, uniformly sampling different ground truth domains (semantic instance, or panoptic) by deriving all labels from panoptic annotations to train its multi-task model.
• OneFormer uses a task token to condition the model on the task in focus, making our architecture task-guided for training, and task-dynamic for inference, all with a single model.

Contents

1. News
2. Installation Instructions
3. Dataset Preparation
4. Execution Instructions
   • Training
   • Evaluation
   • Demo
5. Results
6. Citation

News

• [February 27, 2023]: OneFormer is accepted to CVPR 2023!
• [January 26, 2023]: OneFormer sets new SOTA performance on the the Mapillary Vistas val (both panoptic & semantic segmentation) and Cityscapes test (panoptic segmentation) sets. We’ve released the checkpoints too!
• [January 19, 2023]: OneFormer is now available as a part of the 🤗 HuggingFace transformers library and model hub! 🚀
• [December 26, 2022]: Checkpoints for Swin-L OneFormer and DiNAT-L OneFormer trained on ADE20K with 1280×1280 resolution released!
• [November 23, 2022]: Roboflow cover OneFormer on YouTube! Thanks to @SkalskiP for making the video!
• [November 18, 2022]: Our demo is available on 🤗 Huggingface Space!
• [November 10, 2022]: Project Page, ArXiv Preprint and GitHub Repo are public!
  • OneFormer sets new SOTA on Cityscapes val with single-scale inference on Panoptic Segmentation with 68.5 PQ score and Instance Segmentation with 46.7 AP score!
  • OneFormer sets new SOTA on ADE20K val on Panoptic Segmentation with 51.5 PQ score and on Instance Segmentation with 37.8 AP!
  • OneFormer sets new SOTA on COCO val on Panoptic Segmentation with 58.0 PQ score!

Installation Instructions

• We use Python 3.8, PyTorch 1.10.1 (CUDA 11.3 build).
• We use Detectron2-v0.6.
• For complete installation instructions, please see INSTALL.md.

Dataset Preparation

• We experiment on three major benchmark dataset: ADE20K, Cityscapes and COCO 2017.
• Please see Preparing Datasets for OneFormer for complete instructions for preparing the datasets.

Execution Instructions

Training

• We train all our models using 8 A6000 (48 GB each) GPUs.
• We use 8 A100 (80 GB each) for training Swin-L^† OneFormer and DiNAT-L^† OneFormer on COCO and all models with ConvNeXt-XL^† backbone. We also train the 896x896 models on ADE20K on 8 A100 GPUs.
• Please see Getting Started with OneFormer for training commands.

Evaluation

• Please see Getting Started with OneFormer for evaluation commands.

Demo

• We provide quick to run demos on Colab and Hugging Face Spaces .
• Please see OneFormer Demo for command line instructions on running the demo.

Results

• † denotes the backbones were pretrained on ImageNet-22k.
• Pre-trained models can be downloaded following the instructions given under tools.

ADE20K

| Method | Backbone | Crop Size | PQ | AP | mIoU <br> (s.s) | mIoU <br> (ms+flip) | #params | config | Checkpoint | | :---:| :---: | :---: | :---: | :---:| :---: | :---: | :---: | :---: | :---: | | OneFormer | Swin-L^† | 640×640 | 49.8 | 35.9 | 57.0 | 57.7 | 219M | config | model | | OneFormer | Swin-L^† | 896×896 | 51.1 | 37.6 | 57.4 | 58.3 | 219M | [config](configs/ade20k/swin/onefor