Aligning Pre-training and Fine-tuning in Object Detection [Project Page] [arXiv] [Paper] [Poster]

Official PyTorch Implementation of AlignDet: Aligning Pre-training and Fine-tuning in Object Detection (ICCV 2023)

Existing detection algorithms are constrained by the data, model, and task discrepancies between pre-training and fine-tuning.
AlignDet aligns these discrepancies in an efficient and unsupervised paradigm, leading to significant performance improvements across different settings.

Comparison with other self-supervised pre-training methods on data, models and tasks aspects. AlignDet achieves more efficient, adequate and detection-oriented pre-training

Our pipeline takes full advantage of the existing pre-trained backbones to efficiently pre-train other modules. By incorporating self-supervised pre-trained backbones, we make the first attempt to fully pre-train various detectors using a completely unsupervised paradigm.

Data Download

Please download the COCO 2017 dataset, and the folder structure should be:

data
├── coco
│   ├── annotations
│   ├── filtered_proposals
│   ├── semi_supervised_annotations
│   ├── test2017
│   ├── train2017
│   └── val2017

The folder filtered_proposals for self-supervised pre-training can be downloaded in Google Drive.

The folder semi_supervised_annotations for semi-supervised fine-tuning can be downloaded in Google Drive, or generated by tools/generate_semi_coco.py

Environments

# Sorry our code is not based on latest mmdet 3.0+
pip3 install -r requirements.txt

Pre-training and Fine-tuning Instructions

Pre-training

bash tools/dist_train.sh configs/selfsup/mask_rcnn.py 8 --work-dir work_dirs/selfsup_mask-rcnn

Fine-tuning

Using tools/model_converters/extract_detector_weights.py to extract the weights.

python3 tools/model_converters/extract_detector_weights.py \
work_dirs/selfsup_mask-rcnn/epoch_12.pth  \ # pretrain weights
work_dirs/selfsup_mask-rcnn/final_model.pth  # finetune weights

Fine-tuning models like normal mmdet training process, usually the learning rate is increased by 1.5 times, and the weight decay is reduced to half of the original setting. Please refer to the released logs for more details.

bash tools/dist_train.sh configs/coco/mask_rcnn_r50_fpn_1x_coco.py 8 \
--cfg-options load_from=work_dirs/selfsup_mask-rcnn/final_model.pth \ # load pre-trained weights
optimizer.lr=3e-2 optimizer.weight_decay=5e-5  \ # adjust lr and wd
--work-dir work_dirs/finetune_mask-rcnn_1x_coco_lr3e-2_wd5e-5

Checkpoints and Logs

We show part of the results here, and all the checkpoints & logs can be found in the HuggingFace Space.

Different Methods

Method (Backbone)	Pre-training	Fine-tuning
FCOS (R50)	link	link
RetinaNet (R50)	link	link
Faster R-CNN (R50)	link	link
Mask R-CNN (R50)	link	link
DETR (R50)	link	link
SimMIM (Swin-B)	link	link
CBNet v2 (Swin-L)	link	link

Mask R-CNN with Different Backbones Sizes

Backbone	Pre-training	Fine-tuning
MobileNet v2	link	link
ResNet-18	link	link
Swin-Small	link	link
Swin-Base	link	link

Models with Self-supervised Backbones (ResNet-50)

Mask R-CNN	Pre-training	Fine-tuning
MoCo v2	link	link
PixPro	link	link
SwAV	link	link
RetinaNet	Pre-training	Fine-tuning
MoCo v2	link	link
PixPro	link	link
SwAV	link	link

Transfer Learning on VOC Dataset

Method	Fine-tuning
FCOS	link
RetinaNet	link
Faster R-CNN	link
DETR	link

Citation

If you find our work to be useful for your research, please consider citing.

@InProceedings{AlignDet,
    author    = {Li, Ming and Wu, Jie and Wang, Xionghui and Chen, Chen and Qin, Jie and Xiao, Xuefeng and Wang, Rui and Zheng, Min and Pan, Xin},
    title     = {AlignDet: Aligning Pre-training and Fine-tuning in Object Detection},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2023},
}

liming-ai/AlignDet