H-Deformable-DETR

This is the official implementation of the paper "DETRs with Hybrid Matching".

Authors: Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, Han Hu

News

⛽ ⛽ ⛽ Contact: [email protected]

2023.04.14 Expedit-SAM significantly boosts the inference speed of the ViT-H SAM model by almost 1.5 times. 🍺credits to Weicong Liang🍺

2023.04.11 Swin-L+H-Deformable-DETR + SAM achieves strong COCO instance segmentation results: mask AP=46.8 by simply prompting SAM with our HDETR box predictions. (mask AP=46.5 based on ViTDet) 🍺credits to Zhanhao Liang🍺

2023.03.29 HDETR+R50 based on Detrex achieves better performance: AP=49.1 under 12x training epochs. Thanks to Tianhe Ren🍺

2023.03.22 Expedit-LargeScale-Vision-Transformer (NeurIPS2022) has been open-sourced.

2023.02.28 HDETR has been accepted by CVPR 2023 😉😉😉

2022.11.25 Optimized implementation for hybrid matching is released at pull-request, which parallelizes the matching/loss computations of one2one branch and one2many branch. 🍺credits to Ding Jia🍺

2022.11.17 Code for H-Detic-LVIS is released. 🍺credits to Haodi He🍺

2022.11.10 Code for H-TransTrack is released. 🍺credits to Haojun Yu🍺

2022.10.20 🎉🎉🎉Detrex have supported our H-Deformable-DETR 🍺credits to Ding Jia and Tianhe Ren🍺

2022.09.14 We have supported H-Deformable-DETR w/ ViT-L (MAE), which achieves 56.5 AP on COCO val with 4-scale feature maps without using LSJ (large scale jittering) adopted by the original ViT-Det. We will include the results of H-Deformable-DETR w/ ViT-L (MAE) + LSJ equipped with LSJ soon. 🍺credits to Weicong Liang🍺

2022.09.12 Our H-Deformable-DETR w/ Swin-L achieves 58.2 AP on COCO val with 4-scale feature maps, thus achieving comparable (slightly better) results than the very recent DINO-DETR w/ Swin-L equipped with 4-scale feature maps.

2022.08.31 Code for H-Deformable-DETR-mmdet (support mmdetection2d 🍺credits toYiduo Hao🍺) is released. We will also release the code for H-Mask-Deformable-DETR soon (strong results on both instance segmentation and panoptic segmentation).

Model ZOO

We provide a set of baseline results and trained models available for download:

Models with the ResNet-50 backbone

Name	Backbone	query	epochs	AP	download
Deformable-DETR	R50	300	12	43.7	model
Deformable-DETR	R50	300	36	46.8	model
Deformable-DETR + tricks	R50	300	12	47.0	model
Deformable-DETR + tricks	R50	300	36	49.0	model
H-Deformable-DETR + tricks	R50	300	12	48.7	model
H-Deformable-DETR + tricks	R50	300	36	50.0	model

Models with Swin Transformer backbones

Name	Backbone	query	epochs	AP	download
Deformable-DETR	Swin Tiny	300	12	45.3, 46.0	model, model
Deformable-DETR	Swin Tiny	300	36	49.0,49.6	model, model
Deformable-DETR + tricks	Swin Tiny	300	12	49.3	model
Deformable-DETR + tricks	Swin Tiny	300	36	51.8	model
H-Deformable-DETR + tricks	Swin Tiny	300	12	50.6	model
H-Deformable-DETR + tricks	Swin Tiny	300	36	53.2	model
Deformable-DETR	Swin Large	300	12	51.0	model
Deformable-DETR	Swin Large	300	36	53.7	model
Deformable-DETR + tricks	Swin Large	300	12	54.5	model
Deformable-DETR + tricks	Swin Large	300	36	56.3	model
H-Deformable-DETR + tricks	Swin Large	300	12	55.9	model
H-Deformable-DETR + tricks	Swin Large	300	36	57.1	model
H-Deformable-DETR + tricks	Swin Large	900	12	56.1	model
H-Deformable-DETR + tricks	Swin Large	900	36	57.4	model
H-Deformable-DETR + tricks [topk=300]	Swin Large	900	36	57.6	model

Improving H-Deformable-DETR with weight decay 0.05

Name	Backbone	query	epochs	AP (weight-decay=0.0001)	AP (weight-decay=0.05	download
H-Deformable-DETR + tricks	Swin Tiny	300	12	50.6	51.2	model
H-Deformable-DETR + tricks	Swin Tiny	300	36	53.2	53.7	model
H-Deformable-DETR + tricks	Swin Large	900	36	57.4	57.8	model
H-Deformable-DETR + tricks [topk=300]	Swin Large	900	36	57.6	57.9	model
H-Deformable-DETRdeep-encoder + tricks [topk=300]	Swin Large	900	36	NA	58.2	model

Installation

We test our models under python=3.7.10,pytorch=1.10.1,cuda=10.2. Other versions might be available as well.

1. Clone this repo

git https://github.com/HDETR/H-Deformable-DETR.git
cd H-Deformable-DETR

2. Install Pytorch and torchvision

Follow the instruction on https://pytorch.org/get-started/locally/.

# an example:
conda install -c pytorch pytorch torchvision

3. Install other needed packages

pip install -r requirements.txt
pip install openmim
mim install mmcv-full
pip install mmdet

4. Compiling CUDA operators

cd models/ops
python setup.py build install
# unit test (should see all checking is True)
python test.py
cd ../..

Data

Please download COCO 2017 dataset and organize them as following:

coco_path/
  ├── train2017/
  ├── val2017/
  └── annotations/
  	├── instances_train2017.json
  	└── instances_val2017.json

Run

To train a model using 8 cards

GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 <config path> \
    --coco_path <coco path>

To train/eval a model with the swin transformer backbone, you need to download the backbone from the offical repo frist and specify argument--pretrained_backbone_path like our configs.

To eval a model using 8 cards

GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 <config path> \
    --coco_path <coco path> --eval --resume <checkpoint path>

Distributed Run

You can refer to Deformable-DETR to enable training on multiple nodes.

Modified files compared to vanilla Deformable DETR

To support swin backbones

• models/backbone.py
• models/swin_transformer.py
• mmcv_custom

To support eval in the training set

• datasets/coco.py
• datasets/__init__.py

To support Hybrid-branch, tricks and checkpoint

• main.py
• engine.py
• models/deformable_detr.py
• models/deformable_transformer.py

To support fp16

• models/ops/modules/ms_deform_attn.py
• models/ops/functions/ms_deform_attn_func.py

To fix a pytorch version bug

• util/misc.py

Addictional packages needed

• wandb: for logging
• mmdet: for swin backbones
• mmcv: for swin backbones
• timm: for swin backbones

Citing H-Deformable-DETR

If you find H-Deformable-DETR useful in your research, please consider citing:

@article{jia2022detrs,
  title={DETRs with Hybrid Matching},
  author={Jia, Ding and Yuan, Yuhui and He, Haodi and Wu, Xiaopei and Yu, Haojun and Lin, Weihong and Sun, Lei and Zhang, Chao and Hu, Han},
  journal={arXiv preprint arXiv:2207.13080},
  year={2022}
}

@article{zhu2020deformable,
  title={Deformable detr: Deformable transformers for end-to-end object detection},
  author={Zhu, Xizhou and Su, Weijie and Lu, Lewei and Li, Bin and Wang, Xiaogang and Dai, Jifeng},
  journal={arXiv preprint arXiv:2010.04159},
  year={2020}
}