DETRs with Collaborative Hybrid Assignments Training

[📖 Paper] [🤗 Hugging Face Model]

News

• [07/21/2024] Check out our Co-DETR detection and segmentation checkpoints, fine-tuned on COCO and LVIS, now available on Hugging Face. We've achieved new state-of-the-art performance in instance segmentation!
• [04/22/2024] We release a new MLLM framework MoVA, which adopts Co-DETR as the vision and achieves state-of-the-art performance on multimodal benchmarks.
• [10/19/2023] Our SOTA model Co-DETR w/ ViT-L is released now. Please refer to our huggingface 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.
• [04/05/2023] HoP leverages Co-DETR as the backbone and achieves new SOTA performance on nuScenes 3D detection leaderboard.
• [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

Co-DETR with ResNet-50

Co-DETR with Swin-L

Co-Deformable-DETR

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
    │   │      ├── instances_train2017.json
    │   │      └── instances_val2017.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:

sh tools/dist_train.sh projects/configs/co_deformable_detr/co_deformable_detr_r50_1x_coco.py 8 path_to_exp

Train using slurm:

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:

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:

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.

@inproceedings{zong2023detrs,
  title={Detrs with collaborative hybrid assignments training},
  author={Zong, Zhuofan and Song, Guanglu and Liu, Yu},
  booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
  pages={6748--6758},
  year={2023}
}

License

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