Mind the Gap: Polishing Pseudo labels for Accurate Semi-supervised Object Detection

This repo is the official implementation of our AAAI2023 paper "Mind the Gap: Polishing Pseudo labels for Accurate Semi-supervised Object Detection" on PyTorch.

Results

COCO

1% labeled data

| Method | mAP| Model Weights |Config Files| | :----: | -------| ----- |----| | Baseline| 10.0 |-|Config| | Ours | 23.80 |-|Config|

5% labeled data

| Method | mAP| Model Weights |Config Files| | :----: | -------| ----- |----| | Baseline| 10.00 |-|Config| | Ours | 32.15 |-|Config|

10% labeled data

| Method | mAP| Model Weights |Config Files| | :----: | -------| ----- |----| | Baseline| 20.92 |-|Config| | Ours | 35.30 |-|Config|

VOC

| Method | mAP | AP50 | Model Weights |Config Files| | :----: | -------| ----- |----|----| | Baseline| 43.00 | 76.70 |-|Config| | Ours | 52.40 | 82.50 |-|Config|

Usage

Since this repo is built on the Soft Teacher, some setup instructions are cloned from it.

Requirements

Ubuntu 16.04
Anaconda3 with python=3.6
Pytorch=1.9.0
mmdetection=2.16.0+fe46ffe
mmcv=1.3.9

Installation

pip install -r requirements.txt
cd thirdparty/mmdetection && pip install -e .
cd ../.. && pip install -e .

Data Preparation

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/
#     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

Training

To train model on the partial labeled data setting:

tools/dist_train.sh ./configs/coco/coco_part.py 8 --work-dir=<YOUR_WORKDIR_PATH> --cfg-options fold=<FOLD> percent=<PERCENT_LABELED_DATA>

For example, we could run the following scripts to train our model on 10% labeled data of fold 1 with 8 GPUs:

tools/dist_train.sh ./configs/coco/coco_part.py 8 --work-dir=<YOUR_WORKDIR_PATH> --cfg-options fold=1 percent=10

To train model under VOC07 (as labeled set) and VOC12 (as unlabeled set):

tools/dist_train.sh ./configs/voc/voc07_rf12.py 8 --work-dir=<YOUR_WORKDIR_PATH>

All experiments are trained on 8 GPUs by default.

Evaluation

bash tools/dist_test.sh <CONFIG_FILE_PATH> <CHECKPOINT_PATH> <NUM_GPUS> --eval bbox --cfg-options model.test_cfg.rcnn.score_thr=<THR>

Acknowledgement

Polishing Teacher builds on previous works code base such as Soft Teacher and mmdetection. Thanks for their wonderful works.

Citation

If you found this code helpful, feel free to cite our work:

@article{zhang2022mind,
  title={Mind the Gap: Polishing Pseudo labels for Accurate Semi-supervised Object Detection},
  author={Zhang, Lei and Sun, Yuxuan and Wei, Wei},
  journal={arXiv preprint arXiv:2207.08185},
  year={2022}
}

DualPolishLearning

Install / Use

README