Noisy Boundaries: Lemon or Lemonade for semi-supervised instance segmentation?

This is the mmdetection implementation of our CVPR 2022 paper. ArXiv.

Installation

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

data preparation

noisyboundaries
├──data
|  ├──cityscapes
|  |  ├──annotations
|  |  |  ├──instancesonly_filtered_gtFine_train.json
|  |  |  ├──instancesonly_filtered_gtFine_val.json
|  |  ├──leftImg8bit
|  |  |  ├──train
|  |  |  ├──val
|  ├──coco
|  |  ├──annotations
|  |  |  ├──instances_train2017.json
|  |  |  ├──instances_val2017.json
|  |  ├──images
|  |  |  ├──train2017
|  |  |  ├──val2017

Running scripts

cityscapes

We take the experiment with the 20% labeled images for example.

make the label file first:

mkdir labels
python scripts/cityscapes/prepare_cityscape_data.py --percent 20 --seed 1

Then, to train the supervised model, run:

bash tools/dist_train.sh configs/noisyboundaries/cityscapes/mask_rcnn_r50_fpn_1x_cityscapes_sup.py 8

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

bash scripts/cityscapes/extract_pl.sh 8 labels/rcity.pkl labels/cityscapes_1@20_pl.json 

Then, perform semi-supervised learning:

bash tools/dist_train.sh configs/noisyboundaries/cityscapes/mask_rcnn_r50_fpn_1x_cityscapes_pl.py 8