Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, Heung-Yeung Shum

[arXiv] [BibTeX]

• DINO with modified training engine
• Main Results with Pretrained Models
  • DINO with ResNet Backbone
  • DINO with Swin-Transformer Backbone
  • DINO with ViT Backbone
  • DINO with ConvNeXt Backbone
  • DINO with FocalNet Backbone
  • DINO with InternImage Backbone
  • DINO with EVA Backbone
• Training DINO
• Evaluate DINO
• Citation

We've provide a hacked train_net.py which aligns the optimizer params with Deformable-DETR that can achieve a better result on DINO models.

• Training model with hacked trainer

python projects/dino/train_net.py --config-file /path/to/config.py --num-gpus 8

Note:

• Swin-X-384 means the backbone pretrained resolution is 384 x 384 and IN22k to In1k means the model is pretrained on ImageNet-22k and finetuned on ImageNet-1k.
• ViT backbone using MAE pretraining weights following ViTDet which can be downloaded in MAE. And it's not stable to train ViTDet-DINO without warmup lr-scheduler.
• Focal-LRF-3Level: means using Large-Receptive-Field (LRF) and Focal-Level is setted to 3, please refer to FocalNet for more details about the backbone settings.
• with AMP: means using mixed precision training.
• with EMA: means training with model Exponential Moving Average (EMA).

Notable facts and caveats: The position embedding of DINO in detrex is different from the original repo. We set the tempureture and offsets in PositionEmbeddingSine to 10000 and -0.5 which may make the model converge a little bit faster in the early stage and get a slightly better results (about 0.1mAP) in 12 epochs settings.

All configs can be trained with:

cd detrex
python tools/train_net.py --config-file projects/dino/configs/path/to/config.py --num-gpus 8

By default, we use 8 GPUs with total batch size as 16 for training.

Model evaluation can be done as follows:

cd detrex
python tools/train_net.py --config-file projects/dino/configs/path/to/config.py --eval-only train.init_checkpoint=/path/to/model_checkpoint

If you find our work helpful for your research, please consider citing the following BibTeX entry.

@misc{zhang2022dino,
      title={DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection}, 
      author={Hao Zhang and Feng Li and Shilong Liu and Lei Zhang and Hang Su and Jun Zhu and Lionel M. Ni and Heung-Yeung Shum},
      year={2022},
      eprint={2203.03605},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}