/PoER

Potential energy ranking for domain generalization (DG)

Primary LanguagePython

Domain Decorrelation with Potential Energy Ranking

NICO

NICO

Official PyTorch Implementation

Sen Pei, Jiaxi Sun
Institute of Automation, Chinese Academy of Sciences

Datasets

  • PACS: photo, art painting, cartoon, sketch
  • VLCS: Pascal VOC, LabelMe, Caltech, SUN09
  • OfficeHome: Artistic, Clipart, Product, Real World
  • Digits-DG: MNIST, MNIST-M, SVHN, SYN
  • NICO: 19 classes belonging to 65 domains
  • download here: Baidu Disk (ql17)

Data Augmentation Scheme

  • RandomResizedCrop
  • RandomHorizontalFlip
  • ColorJitter
  • Normalize

Architectures

  • backbone: ResNet-18
  • distance-based cross entropy

Pretrained Models

  • ImageNet Pre-trained models

Training

  • python -m torch.distributed.launch --nproc_per_node=8 pacs_train.py --n_gpus=8
  • python -m torch.distributed.launch --nproc_per_node=8 vlcs_train.py --n_gpus=8
  • python -m torch.distributed.launch --nproc_per_node=8 nico_train.py --n_gpus=8
  • python -m torch.distributed.launch --nproc_per_node=8 digits_train.py --n_gpus=8
  • python -m torch.distributed.launch --nproc_per_node=8 officehome_train.py --n_gpus=8

Metrics

Methods Art Painting Cartoon Photo Sketch Average
MMD-AAE 75.20 72.70 96.00 64.20 77.03
CCSA 80.50 76.90 93.60 66.80 79.45
ResNet-18 77.00 75.90 96.00 69.20 79.53
StableNet 80.16 74.15 94.24 70.10 79.66
JiGen 79.40 75.30 96.00 71.60 80.50
CrossCrad 79.80 76.80 96.00 70.20 80.70
DANN 80.20 77.60 95.40 70.00 80.80
Epi-FCR 82.10 77.00 93.90 73.00 81.50
MetaReg 83.70 77.20 95.50 70.30 81.70
GCPL 82.64 75.02 96.40 73.36 81.86
EISNet 81.89 76.44 95.93 74.33 82.15
L2A-OT 83.30 78.20 96.20 73.60 82.83
MixStyle 84.10 78.80 96.10 75.90 83.70
PoER (Ours) 85.30 77.69 96.42 77.30 84.18

Reference

Citation

@misc{pei2022domain,
      title={Domain Decorrelation with Potential Energy Ranking}, 
      author={Sen Pei and Jiaxi Sun and Richard Yida Xu and and Shiming Xiang and Gaofeng Meng},
      year={2022},
      eprint={2207.12194},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}