rastna12 / fct

Code for CVPR 2022 Oral paper: 'Few-Shot Object Detection with Fully Cross-Transformer'

FCT

This repo contains the official PyTorch implementation for the CVPR 2022 Oral paper: 'Few-Shot Object Detection with Fully Cross-Transformer' (paper).

Highlights

• To the best of our knowledge, we are the first to explore and propose the vision transformer based models for few-shot object detection.
• The proposed FCT model can encourage multi-level interactions between the query and support, gradually from low-level to high-level feature spaces.

Installation

Our codebase is built upon detectron2. You only need to install detectron2 following their instructions.

Please note that we used detectron 0.2.1 in this project. Higher versions of detectron might report errors.

Data Preparation

• We evaluate our model on two FSOD benchmarks PASCAL VOC and MSCOCO following the previous work TFA.
• Please prepare the original PASCAL VOC and MSCOCO datasets and also the few-shot datasets following TFA in the folder ./datasets/coco and ./datasets/pascal_voc respectively.
• Please run the scripts in ./datasets/coco and ./datasets/pascal_voc step by step to generate the support images for both many-shot base classes (used during meta-training) and few-shot classes (used during few-shot fine-tuning).

Converting ImageNet pre-trained PVT models into C4-based detection format

The script is

python build_pvt_C4_from_official_model.py

We use the converted model pvt_v2_b2_li_C4.pth by default next.

Model training and evaluation on MSCOCO

• We have three steps for model training, first pre-training the single-branch based model over base classes, then training the two-branch based model over base classes, and finally fine-tuning the two-branch based model over novel classes.
• The training script for pre-training the single-branch based model over base classes is

sh scripts/single_branch_pretraining_coco_pvt_v2_b2_li.sh

• Then initailized with the first step trained model, the script for training the two-branch based model over base classes is

sh scripts/two_branch_training_coco_pvt_v2_b2_li.sh

• Finally we perform 1/2/3/5/10/30-shot fine-tuning over novel classes, using the exact same few-shot datasets as TFA. The training script is

sh scripts/two_branch_few_shot_finetuning_coco_pvt_v2_b2_li.sh

Model training and evaluation on PASCAL VOC

• We evaluate our model on the three splits as TFA.
• Similar as MSCOCO, we have three steps for model training.
• The training scripts for VOC split1 is

sh scripts/single_branch_pretraining_pascalvoc_split1_pvt_v2_b2_li.sh
sh scripts/two_branch_training_pascalvoc_split1_pvt_v2_b2_li.sh
sh scripts/two_branch_few_shot_finetuning_pascalvoc_split1_pvt_v2_b2_li.sh

• The training scripts for VOC split2 is

sh scripts/single_branch_pretraining_pascalvoc_split2_pvt_v2_b2_li.sh
sh scripts/two_branch_training_pascalvoc_split2_pvt_v2_b2_li.sh
sh scripts/two_branch_few_shot_finetuning_pascalvoc_split2_pvt_v2_b2_li.sh

• The training scripts for VOC split3 is

sh scripts/single_branch_pretraining_pascalvoc_split3_pvt_v2_b2_li.sh
sh scripts/two_branch_training_pascalvoc_split3_pvt_v2_b2_li.sh
sh scripts/two_branch_few_shot_finetuning_pascalvoc_split3_pvt_v2_b2_li.sh

Model Zoo

We provided both the single-branch pre-trained models and the meta-trained models (the two-branch based models) over base classes for both MSCOCO dataset and the 3 splits on VOC dataset. The model links are Google Drive and Tencent Weiyun.

Citing FCT

If you use this work in your research or wish to refer to the baseline results published here, please use the following BibTeX entries:

@inproceedings{han2022few,
  title={Few-shot object detection with fully cross-transformer},
  author={Han, Guangxing and Ma, Jiawei and Huang, Shiyuan and Chen, Long and Chang, Shih-Fu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5321--5330},
  year={2022}
}
@inproceedings{han2022meta,
  title={Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment},
  author={Han, Guangxing and Huang, Shiyuan and Ma, Jiawei and He, Yicheng and Chang, Shih-Fu},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={1},
  pages={780--789},
  year={2022}
}
@inproceedings{han2021query,
  title={Query adaptive few-shot object detection with heterogeneous graph convolutional networks},
  author={Han, Guangxing and He, Yicheng and Huang, Shiyuan and Ma, Jiawei and Chang, Shih-Fu},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={3263--3272},
  year={2021}
}

Acknowledgement

This repo is developed based on Meta Faster R-CNN, QA-FewDet and PVT. Thanks for their wonderful codebases.