Home

Awesome

Deformable Mixer Transformer with Gating for Multi-Task Learning of Dense Prediction

This repo is the official implementation of "DeMTG" as well as the follow-ups. It currently includes code and models for the following tasks:

Updates

07/07/2023 We release the models and code of DeMTG.

Introduction

DeMTG (the name DeMTG stands for Deformable Mixer Transformer with Gating for Multi-Task Learning of Dense Prediction) is initially described in arxiv, which is an extension to our previous AAAI 2023. We introduce deformable mixer Transformer with gating (DeMTG), a simple and effective encoder-decoder architecture up-to-date that incorporates the convolution and attention mechanism in a unified network for MTL. DeMTG achieves strong performance on PASCAL-Context (78.54 mIoU semantic segmentation and 67.42 mIoU Human Segmentation on test) and and NYUD-v2 semantic segmentation (57.55 mIoU on test), surpassing previous models by a large margin.

DeMTG

Performance

DeMTG

Main Results on ImageNet with Pretrained Models

DeMTG on NYUD-v2 dataset

modelbackbone#paramsFLOPsSemSegDepthNoemalBoundarymodel checkpopintlog
DeMTGSwin-T33.2M125.49G47.200.566020.1577.2Google Drivelog
DeMTGSwin-S54.52M145.84G52.230.559920.0578.4Google Drivelog
DeMTGSwin-B94.4M-G54.450.522819.3378.6Google Drivelog
DeMTGSwin-L202.92321.22G57.550.503719.2179.0Google Drivelog

DeMTG on PASCAL-Contex dataset

modelbackboneSemSegPartSegSalNormalBoundarymodel checkpopintlog
DeMTGSwin-T69.4458.0283.3114.3171.2Google Drivelog
DeMTGSwin-S71.5461.4983.7014.9072.2Google Drivelog
DeMTGSwin-B75.3764.8283.7514.2273.0Google Drivelog
DeMTGSwin-L78.5467.4283.7414.1774.9Google Drivelog

Citation

@inproceedings{xyy2023DeMT,
  title={DeMT: Deformable Mixer Transformer for Multi-Task Learning of Dense Prediction},
  author={Xu, Yangyang and Yang, Yibo and Zhang, Lefei },
  booktitle={Proceedings of the The Thirty-Seventh Conference on Artificial Intelligence (AAAI)},
  year={2023}
}

@inproceedings{xyy2023DeMTG,
  title={Deformable Mixer Transformer with Gating for Multi-Task Learning of Dense Prediction},
  author={Xu, Yangyang and Yang, Yibo and Ghanem, Bernard and Zhang, Lefei and Bo, Du and Tao, Dacheng},
  booktitle={arxiv},
  year={2023}
}

Getting Started

Install and Data Prepare

Please reference to DeMT

Train

To train DeMTG model:

python ./src/main.py --cfg ./config/t-nyud/swin/siwn_t_DeMTG.yaml --datamodule.data_dir $DATA_DIR --trainer.gpus 8

Evaluation

Acknowledgement

This repository is based ATRC. Thanks to ATRC!