This repository contains the offical implementation for our CVPR-2023 paper.

✨We are now able to train detector on 10% MS-COCO to 40 mAP✨

Consistent-Teacher: Towards Reducing Inconsistent Pseudo-targets in Semi-supervised Object Detection

[arxiv] [code] [project page]

Xinjiang Wang*, Xingyi Yang*, Shilong Zhang, Yijiang Li, Litong Feng, Shijie Fang, Chengqi Lyu, Kai Chen, Wayne Zhang

(*: Co-first Author)

• Selected as Hightligh for CVPR2023🔥 (235/2360, top 10% accepted paper)

In this paper, we systematically investigate the inconsistency problems in semi-supervised object detection, where the pseudo boxes may be highly inaccurate and vary greatly at different stages of training. To alleviate the aforementioned problem, we present a holistic semi-supervised object detector termed Consistent-Teacher. Consistent-Teacher achieves compelling improvement on a wide range of evaluations and serves as a new solid baseline for SSOD.

All results, logs, configs and checkpoints are listed here. Enjoy 👀!

MS-COCO 1%/2%/5/%/10% Labeled Data

MS-COCO100% Labeled + Unlabeled Data

PASCAL VOC07 Label + VOC12 Unlabel

• Defaultly, all models are trained on 8*V100 GPUs with 5 images per GPU.
• Additionally, we support the 2x8 and fp16 training setting to ensure everyone is able to run the code, even with only 12G graphic cards.
• With 8x2+fp16, the total training time for MS-COCO is less than 1 day.
• We carefully tuned the hyper-parameters after submitting the paper, which is why the results in the repository are slightly higher than those reported in the paper.

Zoom in for better View.

├── configs              
    ├── baseline
    │   |-- mean_teacher_retinanet_r50_fpn_coco_180k_10p.py       
    |       # Mean Teacher COCO 10% config
    |   |-- mean_teacher_retinanet_r50_fpn_voc0712_72k.py      
    |       # Mean Teacher VOC0712 config
    ├── consistent-teacher
    |   |-- consistent_teacher_r50_fpn_coco_360k_fulldata.py           
    |       # Consistent Teacher COCO label+unlabel config
    |
    |   |-- consistent_teacher_r50_fpn_coco_180k_1/2/5/10p.py           
    |       # Consistent Teacher COCO 1%/2%/5%/10% config
    |   |-- consistent_teacher_r50_fpn_coco_180k_10p_2x8.py     
    |       # Consistent Teacher COCO 10% config with 8x2 GPU
    |   |-- consistent_teacher_r50_fpn_voc0712_72k.py             
    |       # Consistent Teacher VOC0712 config
├── ssod
    |-- models/mean_teacher.py                           
    |   # Consistent Teacher Class file
    |-- models/consistent_teacher.py                     
    |   # Consistent Teacher Class file
    |-- models/dense_heads/fam3d.py                      
    |   # FAM-3D Class file
    |-- models/dense_heads/improved_retinanet.py                      
    |   # ImprovedRetinaNet baseline file
    |-- core/bbox/assigners/dynamic_assigner.py
    |   # Aadaptive Sample Assignment Class file
├── tools
    |-- dataset/semi_coco.py
    |   # COCO data preprocessing
    |-- train.py/test.py
    |   # Main file for train and evaluate the models

• Pytorch=1.9.0
• mmdetection=2.25.0
• mmcv=1.3.9
• wandb=0.10.31

or

• mmdetection=2.28.1
• mmcv=1.7.1

• We use wandb for visualization, if you don't want to use it, just comment line 328-339 in configs/consistent-teacher/consistent_teacher_r50_fpn_coco_180k_10p.py.

Install all the requirements INSTALL, then git pull the mmdetecton repo and ConsistentTeacher under the same folder

git clone https://github.com/open-mmlab/mmdetection.git
git clone https://github.com/Adamdad/ConsistentTeacher.git
cd ConsistentTeacher/
make install

• Download the COCO dataset
• Execute the following command to generate data set splits:

# YOUR_DATA should be a directory contains coco dataset.
# For eg.:
# YOUR_DATA/
#  coco_semi/
#     instances_train2017.${fold}@${percent}.json
#  coco/
#     train2017/
#     val2017/
#     unlabeled2017/
#     annotations/
ln -s ${YOUR_DATA} data
bash tools/dataset/prepare_coco_data.sh conduct

For concrete instructions of what should be downloaded, please refer to tools/dataset/prepare_coco_data.sh line 11-24

• Download JSON files for unlabeled images PASCAL VOC data in COCO format

cd ${DATAROOT}

wget https://storage.cloud.google.com/gresearch/ssl_detection/STAC_JSON.tar
tar -xf STAC_JSON.tar.gz
# voc/VOCdevkit/VOC2007/instances_test.json
# voc/VOCdevkit/VOC2007/instances_trainval.json
# voc/VOCdevkit/VOC2012/instances_trainval.json

• To train model on the partial labeled data and full labeled data setting:

# CONFIG_FILE_PATH: the config file for experiment.
# GPU_NUM: number of gpus to run the job
bash tools/dist_train.sh <CONFIG_FILE_PATH> <NUM_GPUS>

For example, to train ours R50 model with 8 GPUs:

bash tools/dist_train.sh configs/consistent-teacher/consistent_teacher_r50_fpn_coco_180k_10p.py 8

• To train model on new dataset:

The core idea is to convert a new dataset to coco format. Details about it can be found in the adding new dataset.

• To inference with the pretrained models on images and videos and plot the bounding boxes, we add two scripts
  • tools/inference.py for image inference
  • tools/inference_vido.py for video inference

This project is released under the Apache 2.0 license.

@article{wang2023consistent,
    author    = {Xinjiang Wang, Xingyi Yang, Shilong Zhang, Yijiang Li, Litong Feng, Shijie Fang, Chengqi Lyu, Kai Chen, Wayne Zhang },
    title     = {Consistent-Teacher: Towards Reducing Inconsistent Pseudo-targets in Semi-supervised Object Detection},
    journal   = {The IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR)},
    year      = {2023},
}

• This code pattern was inspired from a SoftTeacher and mmdet