DETRs with Collaborative Hybrid Assignments Training

This repo is the official implementation of "DETRs with Collaborative Hybrid Assignments Training" by Zhuofan Zong, Guanglu Song, and Yu Liu.

News

• [10/19/2023] Our SOTA model Co-DETR w/ ViT-L is released now. Please refer to this page for more details.
• [09/10/2023] We release LVIS inference configs and a stronger LVIS detector that achieves 64.5 box AP.
• [08/21/2023] Our O365 pre-trained Co-DETR with Swin-L achieves 64.8 AP on COCO test-dev. The config and weights are released.
• [07/20/2023] Code for Co-DINO is released: 55.4 AP with ResNet-50 and 60.7 AP with Swin-L.
• [07/14/2023] Co-DETR is accepted to ICCV 2023!
• [07/12/2023] We finetune Co-DETR on LVIS and achieve the best results without TTA: 72.0 box AP and 59.7 mask AP on LVIS minival, 68.0 box AP and 56.0 mask AP on LVIS val. For instance segmentation, we report the performance of the auxiliary mask branch.
• [07/03/2023] Co-DETR with ViT-L (304M parameters) sets a new record of 65.6 66.0 AP on COCO test-dev, surpassing the previous best model InternImage-G (~3000M parameters). It is the first model to exceed 66.0 AP on COCO test-dev.
• [07/03/2023] Code for Co-Deformable-DETR is released.
• [11/19/2022] We achieved 64.4 AP on COCO minival and 64.5 AP on COCO test-dev with only ImageNet-1K as pre-training data. Codes will be available soon.

Introduction

In this paper, we present a novel collaborative hybrid assignments training scheme, namely Co-DETR, to learn more efficient and effective DETR-based detectors from versatile label assignment manners. 1. Encoder optimization: The proposed training scheme can easily enhance the encoder's learning ability in end-to-end detectors by training multiple parallel auxiliary heads supervised by one-to-many label assignments. 2. Decoder optimization: We conduct extra customized positive queries by extracting the positive coordinates from these auxiliary heads to improve attention learning of the decoder. 3. State-of-the-art performance: Co-DETR with ViT-L (304M parameters) is the first model to achieve 66.0 AP on COCO test-dev.

Model Zoo

Objects365 pre-trained Co-DETR

Note: the inconsistent pre-training and fine-tuning augmentation settings (DETR and LSJ aug) for the Swin-L model degenerate the performance on LVIS. | Model | Backbone | Epochs | Aug | Dataset | box AP (val) | Config | Download | | ------ | -------- | ------ | --- | ------- | ------------ | ------ | ----- | | Co-DINO | Swin-L | 16 | DETR | COCO | 64.1 | config | model | | Co-DINO | Swin-L | 16 | LSJ | LVIS | 64.5 | config (test) | model | | Co-DINO | ViT-L | 16 | LSJ | LVIS | 68.0 | config (test) | model |

Co-DETR with ResNet-50

| Model | Backbone | Epochs | Aug | Dataset | box AP | Config | Download | | ------ | -------- | ------ | --- | ------- | ------ | ------ | ----- | | Co-DINO | R50 | 12 | DETR | COCO | 52.1 | config | model | | Co-DINO | R50 | 12 | LSJ | COCO | 52.1 | config | model | | Co-DINO-9enc | R50 | 12 | LSJ | COCO | 52.6 | config | model | | Co-DINO | R50 | 36 | LSJ | COCO | 54.8 | config | model | | Co-DINO-9enc | R50 | 36 | LSJ | COCO | 55.4 | config | model |

Co-DETR with Swin-L

| Model | Backbone | Epochs | Aug | Dataset | box AP | Config | Download | | ------ | -------- | ------ | --- | ------- | ------ | ------ | ----- | | Co-DINO | Swin-L | 12 | DETR | COCO | 58.9 | config | model | | Co-DINO | Swin-L | 24 | DETR | COCO | 59.8 | config | model | | Co-DINO | Swin-L | 36 | DETR | COCO | 60.0 | config | model | | Co-DINO | Swin-L | 12 | LSJ | COCO | 59.3 | config | model | | Co-DINO | Swin-L | 24 | LSJ | COCO | 60.4 | config | model | | Co-DINO | Swin-L | 36 | LSJ | COCO | 60.7 | config | model | | Co-DINO | Swin-L | 36 | LSJ | LVIS | 56.9 | config (test) | model |

Co-Deformable-DETR

| Model | Backbone | Epochs | Queries | box AP | Config | Download | | ------ | -------- | ------ | ------- | ------ | ---- | --- | | Co-Deformable-DETR | R50 | 12 | 300 | 49.5 | config | model | log | | Co-Deformable-DETR | Swin-T | 12 | 300 | 51.7 | config | model | log | | Co-Deformable-DETR | Swin-T | 36 | 300 | 54.1 | config | model | log | | Co-Deformable-DETR | Swin-S | 12 | 300 | 53.4 | config | model | log | | Co-Deformable-DETR | Swin-S | 36 | 300 | 55.3 | config | model | log | | Co-Deformable-DETR | Swin-B | 12 | 300 | 55.5 | config | model | log | | Co-Deformable-DETR | Swin-B | 36 | 300 | 57.5 | config | model | log | | Co-Deformable-DETR | Swin-L | 12 | 300 | 56.9 | config | model | log | | Co-Deformable-DETR | Swin-L | 36 | 900 | 58.5 | config | model | log |

Running

Install

We implement Co-DETR using MMDetection V2.25.3 and MMCV V1.5.0. The source code of MMdetection has been included in this repo and you only need to build MMCV following official instructions. We test our models under python=3.7.11,pytorch=1.11.0,cuda=11.3. Other versions may not be compatible.

Data

The COCO dataset and LVIS dataset should be organized as: ``` Co-DETR data coco annotations instancestrain2017.json instancesval2017.json train2017 val2017

 lvis_v1
     annotations
           lvis_v1_train.json
           lvis_v1_val.json
     train2017
     val2017        

```

Training

Train Co-Deformable-DETR + ResNet-50 with 8 GPUs: shell sh tools/dist_train.sh projects/configs/co_deformable_detr/co_deformable_detr_r50_1x_coco.py 8 path_to_exp Train using slurm: shell sh tools/slurm_train.sh partition job_name projects/configs/co_deformable_detr/co_deformable_detr_r50_1x_coco.py path_to_exp

Testing

Test Co-Deformable-DETR + ResNet-50 with 8 GPUs, and evaluate: shell sh tools/dist_test.sh projects/configs/co_deformable_detr/co_deformable_detr_r50_1x_coco.py path_to_checkpoint 8 --eval bbox Test using slurm: shell sh tools/slurm_test.sh partition job_name projects/configs/co_deformable_detr/co_deformable_detr_r50_1x_coco.py path_to_checkpoint --eval bbox

Cite Co-DETR

If you find this repository useful, please use the following BibTeX entry for citation.

latex @misc{codetr2022, title={DETRs with Collaborative Hybrid Assignments Training}, author={Zhuofan Zong and Guanglu Song and Yu Liu}, year={2022}, eprint={2211.12860}, archivePrefix={arXiv}, primaryClass={cs.CV} }

License

This project is released under the MIT license. Please see the LICENSE file for more information.