This is the official implementation of the paper "DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection". (DINO pronounced `daɪnoʊ' as in dinosaur)
Authors: Hao Zhang*, Feng Li*, Shilong Liu*, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, Heung-Yeung Shum
[2023/4/28]: We release a strong open-set object detection and segmentation model OpenSeeD that achieves the best results on open-set object segmentation tasks. Code and checkpoints are available here.
[2023/4/22]: With better hyper-params, our DINO-4scale model achieves 49.7AP under 12ep settings, please check detrex: DINO for more details.
[2023/3/13]: We release a strong open-set object detection model Grounding DINO that achieves the best results on open-set object detection tasks. It achieves 52.5 zero-shot AP on COCO detection, without any COCO training data! It achieves 63.0 AP on COCO after fine-tuning. Code and checkpoints will be available here.
[2023/1/23]: DINO has been accepted to ICLR 2023!
[2022/12/02]: Code for Mask DINO is released (also in detrex)! Mask DINO further Achieves 51.7 and 59.0 box AP on COCO with a ResNet-50 and SwinL without extra detection data, outperforming DINO under the same setting!.
[2022/9/22]: We release a toolbox detrex that provides state-of-the-art Transformer-based detection algorithms. It includes DINO with better performance. Welcome to use it!
- Supports Now: DETR, Deformble DETR, Conditional DETR, DAB-DETR, DN-DETR, DINO.
[2022/9/18]: We organize ECCV Workshop Computer Vision in the Wild (CVinW), where two challenges are hosted to evaluate the zero-shot, few-shot and full-shot performance of pre-trained vision models in downstream tasks:
- ``Image Classification in the Wild (ICinW)'' Challenge evaluates on 20 image classification tasks.
- ``Object Detection in the Wild (ODinW)'' Challenge evaluates on 35 object detection tasks.
[2022/8/6]: We update Swin-L model results without techniques such as O365 pre-training, large image size, and multi-scale test. We also upload the corresponding checkpoints to Google Drive. Our 5-scale model without any tricks obtains 58.5 AP on COCO val.
[2022/7/14]: We release the code with Swin-L and Convnext backbone.
[2022/7/10]: We release the code and checkpoints with Resnet-50 backbone.
[2022/6/7]: We release a unified detection and segmentation model Mask DINO that achieves the best results on all the three segmentation tasks (54.7 AP on COCO instance leaderboard, 59.5 PQ on COCO panoptic leaderboard, and 60.8 mIoU on ADE20K semantic leaderboard)! Code will be available here.
[2022/5/28] Code for DN-DETR is available here.
[2020/4/10]: Code for DAB-DETR is avaliable here.
[2022/3/8]: We reach the SOTA on MS-COCO leader board with 63.3AP!
[2022/3/9]: We build a repo Awesome Detection Transformer to present papers about transformer for detection and segmenttion. Welcome to your attention!
We present DINO (DETR with Improved deNoising anchOr boxes) with:
- State-of-the-art & end-to-end: DINO achieves 63.2 AP on COCO Val and 63.3 AP on COCO test-dev with more than ten times smaller model size and data size than previous best models.
- Fast-converging: With the ResNet-50 backbone, DINO with 5 scales achieves 49.4 AP in 12 epochs and 51.3 AP in 24 epochs. Our 4-scale model achieves similar performance and runs at 23 FPS.
We have put our model checkpoints here [model zoo in Google Drive][model zoo in 百度网盘]（提取码"DINO"), where checkpoint{x}_{y}scale.pth denotes the checkpoint of y-scale model trained for x epochs.
| name | backbone | box AP | Checkpoint | Where in Our Paper | |
|---|---|---|---|---|---|
| 1 | DINO-4scale | R50 | 49.0 | Google Drive / BaiDu | Table 1 |
| 2 | DINO-5scale | R50 | 49.4 | Google Drive / BaiDu | Table 1 |
| 3 | DINO-4scale | Swin-L | 56.8 | Google Drive | |
| 4 | DINO-5scale | Swin-L | 57.3 | Google Drive |
| name | backbone | box AP | Checkpoint | Where in Our Paper | |
|---|---|---|---|---|---|
| 1 | DINO-4scale | R50 | 50.4 | Google Drive / BaiDu | Table 2 |
| 2 | DINO-5scale | R50 | 51.3 | Google Drive / BaiDu | Table 2 |
| name | backbone | box AP | Checkpoint | Where in Our Paper | |
|---|---|---|---|---|---|
| 1 | DINO-4scale | R50 | 50.9 | Google Drive / BaiDu | Table 2 |
| 2 | DINO-5scale | R50 | 51.2 | Google Drive / BaiDu | Table 2 |
| 3 | DINO-4scale | Swin-L | 58.0 | Google Drive | |
| 4 | DINO-5scale | Swin-L | 58.5 | Google Drive |
Installation
We use the environment same to DAB-DETR and DN-DETR to run DINO. If you have run DN-DETR or DAB-DETR, you can skip this step.
We test our models under python=3.7.3,pytorch=1.9.0,cuda=11.1. Other versions might be available as well. Click the Details below for more details.
- Clone this repo
git clone https://github.com/IDEA-Research/DINO.git
cd DINO- Install Pytorch and torchvision
Follow the instruction on https://pytorch.org/get-started/locally/.
# an example:
conda install -c pytorch pytorch torchvision- Install other needed packages
pip install -r requirements.txt- Compiling CUDA operators
cd models/dino/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:
COCODIR/
├── train2017/
├── val2017/
└── annotations/
├── instances_train2017.json
└── instances_val2017.json
1. Eval our pretrianed models
Download our DINO model checkpoint "checkpoint0011_4scale.pth" from this link and perform the command below. You can expect to get the final AP about 49.0.
bash scripts/DINO_eval.sh /path/to/your/COCODIR /path/to/your/checkpoint2. Inference and Visualizations
For inference and visualizations, we provide a notebook as an example.
3. Train a 4-scale model for 12 epochs
We use the DINO 4-scale model trained for 12 epochs as an example to demonstrate how to evaluate and train our model.
You can also train our model on a single process:
bash scripts/DINO_train.sh /path/to/your/COCODIR4. Supports for Swin Transformer
To train Swin-L model, you need to first download the checkpoint of Swin-L backbone from link and specify the dir of the pre-trained backbone when running the scripts. Here is an example.
bash scripts/DINO_train_submitit_swin.sh /path/to/your/COCODIR /path/to/your/pretrained_backbone
5. Distributed Run
As the training is time consuming, we suggest to train the model on multi-device.
If you plan to train the models on a cluster with Slurm, here is an example command for training:
# for DINO-4scale: 49.0
bash scripts/DINO_train_submitit.sh /path/to/your/COCODIR
# for DINO-5scale: 49.4
bash scripts/DINO_train_submitit_5scale.sh /path/to/your/COCODIRNotes: The results are sensitive to the batch size. We use 16(2 images each GPU x 8 GPUs for DINO-4scale and 1 images each GPU x 16 GPUs for DINO-5scale) by default.
Or run with multi-processes on a single node:
# for DINO-4scale: 49.0
bash scripts/DINO_train_dist.sh /path/to/your/COCODIR6. Training/Fine-tuning a DINO on your custom dataset
To train a DINO on a custom dataset from scratch, you need to tune two parameters in a config file:
- Tuning the
num_classesto the number of classes to detect in your dataset. - Tuning the parameter
dn_labebook_sizeto ensure thatdn_labebook_size >= num_classes + 1
To leverage our pre-trained models for model fine-tuning, we suggest add two more commands in a bash:
--pretrain_model_path /path/to/a/pretrianed/model. specify a pre-trained model.--finetune_ignore label_enc.weight class_embed. ignore some inconsistent parameters.
Our model is based on DAB-DETR and DN-DETR.
DN-DETR: Accelerate DETR Training by Introducing Query DeNoising.
Feng Li*, Hao Zhang*, Shilong Liu, Jian Guo, Lionel M. Ni, Lei Zhang.
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2022.
[paper] [code] [中文解读]
DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR.
Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, Lei Zhang.
International Conference on Learning Representations (ICLR) 2022.
[paper] [code]
We also thank great previous work including DETR, Deformable DETR, SMCA, Conditional DETR, Anchor DETR, Dynamic DETR, etc. More related work are available at Awesome Detection Transformer.
DINO is released under the Apache 2.0 license. Please see the LICENSE file for more information.
Copyright (c) IDEA. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use these files except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
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}
}
@inproceedings{li2022dn,
title={Dn-detr: Accelerate detr training by introducing query denoising},
author={Li, Feng and Zhang, Hao and Liu, Shilong and Guo, Jian and Ni, Lionel M and Zhang, Lei},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={13619--13627},
year={2022}
}
@inproceedings{
liu2022dabdetr,
title={{DAB}-{DETR}: Dynamic Anchor Boxes are Better Queries for {DETR}},
author={Shilong Liu and Feng Li and Hao Zhang and Xiao Yang and Xianbiao Qi and Hang Su and Jun Zhu and Lei Zhang},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=oMI9PjOb9Jl}
}