This is the PyTorch implementation of our paper:
Unbiased Teacher for Semi-Supervised Object Detection
Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, Peter Vajda
International Conference on Learning Representations (ICLR), 2021

[arXiv] [OpenReview] [Project]

• Linux or macOS with Python ≥ 3.6
• PyTorch ≥ 1.5 and torchvision that matches the PyTorch installation.

# create conda env
conda create -n detectron2 python=3.6
# activate the enviorment
conda activate detectron2
# install PyTorch >=1.5 with GPU
conda install pytorch torchvision -c pytorch

Follow the INSTALL.md to install Detectron2.

1. Download COCO dataset

# download images
wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip

# download annotations
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip

2. Organize the dataset as following:

unbiased_teacher/
└── datasets/
    └── coco/
        ├── train2017/
        ├── val2017/
        └── annotations/
        	├── instances_train2017.json
        	└── instances_val2017.json

• Train the Unbiased Teacher under 1% COCO-supervision

python train_net.py \
      --num-gpus 8 \
      --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup1_run1.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16

• Train the Unbiased Teacher under 2% COCO-supervision

python train_net.py \
      --num-gpus 8 \
      --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup2_run1.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16

• Train the Unbiased Teacher under 5% COCO-supervision

python train_net.py \
      --num-gpus 8 \
      --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup5_run1.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16

• Train the Unbiased Teacher under 10% COCO-supervision

python train_net.py \
      --num-gpus 8 \
      --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16

• Train the Unbiased Teacher under VOC07 (as labeled set) and VOC12 (as unlabeled set)

python train_net.py \
      --num-gpus 8 \
      --config configs/voc/voc07_voc12.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 8 SOLVER.IMG_PER_BATCH_UNLABEL 8

• Train the Unbiased Teacher under VOC07 (as labeled set) and VOC12 COCO20cls (as unlabeled set)

python train_net.py \
      --num-gpus 8 \
      --config configs/voc/voc07_voc12coco20.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 8 SOLVER.IMG_PER_BATCH_UNLABEL 8

python train_net.py \
      --resume \
      --num-gpus 8 \
      --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16 MODEL.WEIGHTS <your weight>.pth

python train_net.py \
      --eval-only \
      --num-gpus 8 \
      --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1.yaml \
       SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16 MODEL.WEIGHTS <your weight>.pth

For the following COCO-supervision results, we use 16 labeled images 16 unlabeled images on 8 GPUs (single node).

Faster-RCNN:

For the following VOC results, we use 8 labeled images 8 unlabeled images on 4 GPUs (single node).

VOC:

1. Q: Using the lower batch size and fewer GPUs cannot achieve the results presented in the paper?

• A: We train the model with 32 labeled images 32 unlabeled images per batch for the results presented in the paper, and using the lower batch size leads to lower accuracy. For example, in the 1% COCO-supervision setting, the model trained with 16 labeled images 16 unlabeled images achieves 20.16 AP as shown in the following table.

1. Q: How to use customized dataset other than COCO and VOC?

• A: Check issue #10. Vladimir Fomenko provides a great answer!

3. Q: What is COCO_supervision.txt? Could I remove it if I need to use my own dataset?

• A: COCO_supervision.txt stores data split of the results we presented in the paper. We did this to make sure the results are reproducible. Also, we found out that the variance across runs is less than 1 mAP, so using other random seed should lead to similar results.

4. Why VOC results in github repo look better than VOC results presented in the paper?

• A: We use COCOevaluator to evalute VOC07-test on paper, while we notice that VOCevaluator has different way to compute AP and results in higher results.

If you use Unbiased Teacher in your research or wish to refer to the results published in the paper, please use the following BibTeX entry.

@inproceedings{liu2021unbiased,
    title={Unbiased Teacher for Semi-Supervised Object Detection},
    author={Liu, Yen-Cheng and Ma, Chih-Yao and He, Zijian and Kuo, Chia-Wen and Chen, Kan and Zhang, Peizhao and Wu, Bichen and Kira, Zsolt and Vajda, Peter},
    booktitle={Proceedings of the International Conference on Learning Representations (ICLR)},
    year={2021},
}

Also, if you use Detectron2 in your research, please use the following BibTeX entry.

@misc{wu2019detectron2,
  author =       {Yuxin Wu and Alexander Kirillov and Francisco Massa and
                  Wan-Yen Lo and Ross Girshick},
  title =        {Detectron2},
  howpublished = {\url{https://github.com/facebookresearch/detectron2}},
  year =         {2019}
}

This project is licensed under MIT License, as found in the LICENSE file.