Combating Noise: Semi-supervised Learning by Region Uncertainty Quantification

This is the mmdetection implementation of our NeurIPS 2021 paper:

Zhenyu Wang, Yali Li, Ye Guo, Shengjin Wang. Combating Noise: Semi-supervised Learning by Region Uncertainty Quantification. ArXiv.

Installation

This code is based on mmdetection v2.18. Please install the code according to the mmdetection step first.

data preparation

multiphase
├──data
|  ├──VOCdevkit
|  |  ├──VOC2007
|  |  ├──VOC2012
|  ├──coco
|  |  ├──annotations
|  |  |  ├──instances_train2014.json
|  |  |  ├──instances_valminusminival2014.json
|  |  |  ├──instances_minival2014.json
|  |  ├──images
|  |  |  ├──train2014
|  |  |  ├──val2014

Running scripts

pascal voc

Run:

python tools/dataset_converters/pascal_voc.py data/VOCdevkit -o labels --out-format coco
python scripts/addscore.py labels/voc07_trainval.json

to prepare the dataset.

Then, to train the supervised model, run (the default gpu number for VOC is 4):

bash tools/dist_train.sh configs/combatnoise/pascal_voc/faster_rcnn_r50_fpn_1x_voc07_sup.py 4

With the supervised model, generating pseudo labels for semi-supervised learning:

bash scripts/pascal_voc/extract_pl.sh 4 labels/rvoc.pkl labels/voc12_trainval_pl.json

Then, perform semi-supervised learning:

bash tools/dist_train.sh configs/combatnoise/pascal_voc/faster_rcnn_r50_fpn_1x_voc07_pl.py 4

coco

python scripts/addscore.py data/coco/annotations/instances_valminusminival2014.json
bash tools/dist_train.sh configs/combatnoise/coco/faster_rcnn_r50_fpn_1x_coco_sup.py 8
bash scripts/coco/extract_pl.sh 8 labels/rcoco.pkl labels/cocotrain2014_pl.json 
bash tools/dist_train.sh configs/combatnoise/coco/faster_rcnn_r50_fpn_1x_coco_pl.py 8

Future features

Experiments on COCO partial (1%, 2%, 5%, 10% ratio for labeled images)