partial-labels

Presentation

This is the working repository for a research project about partial labels. Authors are Laura Calem and Olivier Petit

Setup

1. Download repo

Go in some folder <repo_root>
git clone git@github.com:lcalem/partial-labels.git .
Put <repo_root>/partial-labels in your PYTHONPATH by putting this line in your /.bashrc or whatever file you're using:

export PYTHONPATH="${PYTHONPATH}:/<repo_root>/partial-labels"

2. Datasets

2.1. PascalVOC

2.1.1. Download
Make a dataset folder and cd in it (it will be called <dataset_root>)
Download Pascal-VOC 2007 dataset

trainval wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
test wget wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar

2.1.2. Untar everything in place and all annotations / images should come in place nicely

tar xvf VOCtrainval_06-Nov-2007.tar
tar xvf VOCtest_06-Nov-2007.tar

2.1.3. Preprocess using data/pascalvoc/preprocessing/pp_multilabel.py to create a single annotation csv (do it once for trainval and a second time for test, separately)

python3 pp_multilabel.py <dataset_root> train
python3 pp_multilabel.py <dataset_root> val

2.1.4. Use data/pascalvoc/preprocessing/partial_datasets.py to create partial datasets:

python3 partial_datasets.py <dataset_root>/Annotations

2.2. MS COCO

2.2.1. Download
Make a dataset folder and cd in it (it will be called <dataset_root>)
Download MSCOCO 2014 dataset

train images wget http://images.cocodataset.org/zips/train2014.zip
val images wget http://images.cocodataset.org/zips/val2014.zip
train+val annotations wget http://images.cocodataset.org/annotations/annotations_trainval2014.zip

2.2.2. Unzip everything in place

2.2.3. Preprocess using data/coco/preprocessing/pp_multilabel.py to create a single annotation csv (do it once for train and once for val separately):

python3 pp_multilabel.py <dataset_root> train2014
python3 pp_multilabel.py <dataset_root> val2014

This operation will create csv complete datasets <dataset_root>/annotations/multilabel_train2014.csv and <dataset_root>/annotations/multilabel_val2014.csv.

2.2.4. Use data/coco/preprocessing/partial_datasets.py to create partial datasets (one with 10% known labels, one with 20% known labels, and so on til 100% known labels which should be identical to multilabel_train2014.csv):

python3 partial_datasets.py <dataset_root>/annotations/multilabel_train2014.csv