This repo contains the code for reproducing object detection results of our FocalNets. It is based on DINO.
Please follow DINO's instruction for installation.
- Train on COCO with FocalNet-L with 3 focal levels:
python -m torch.distributed.launch --nproc_per_node={ngpus} main.py --config_file config/DINO/DINO_4scale_focalnet_fl3.py --coco_path {coco_path} --output_dir {output_dir}
- Train on COCO with 5scale DINO and FocalNet-L with 4 focal levels:
python -m torch.distributed.launch --nproc_per_node={ngpus} main.py --config_file config/DINO/DINO_5scale_focalnet_fl4.py --coco_path {coco_path} --output_dir {output_dir}
FocalNet-DINO pretrained with Object365:
| Backbone | Method | Pretrained Data | COCO minival mAP (w/o tta) | Download |
|---|---|---|---|---|
| Swin-L | DINO | Object365 | 63.1 | - |
| FocalNet-L | DINO | Object365 | 63.5 | in21k ckpt/o365 ckpt/coco ckpt |
Thanks to the authors of DINO, the DINO models trained with FocalNets as the backbones can be found here:
FocalNet-L + DINO: DINO + FocalNet-L
All pretrained models on imagenet-1k or imagenet-21k are provided in:
Focal Modulation Networks: Focal Modulation Networks Model Zoo.
If you find this repo useful to your project, please consider to cite it with following bib:
@misc{yang2022focalnet,
author = {Yang, Jianwei and Li, Chunyuan and Dai, Xiyang and Yuan, Lu and Gao, Jianfeng},
title = {Focal Modulation Networks},
publisher = {arXiv},
year = {2022},
}
and also:
@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}
}