Data Preparation

October 12, 2025 · View on GitHub

Our setup follows Matcher.
Create a new directory data to store all the datasets. At the end of the preparation, we expect the following directory structure:

SANSA/
├── data/
│   ├── COCO2014/           
│   │   ├── annotations/
│   │   │   ├── train2014/
│   │   │   └── val2014/
│   │   ├── train2014/
│   │   ├── val2014/
│   │   └── splits/
│   │       ├── trn/
│   │       └── val/
│   ├── FSS-1000/
│   │   ├── data/
│   │   │   ├── ab_wheel/
│   │   │   ├── ...
│   │   │   └── zucchini/
│   │   └── splits/
│   │       ├── trn.txt
│   │       ├── val.txt
│   │       └── test.txt
│   ├── LVIS/
│   │   ├── coco/
│   │   │   ├── train2017/
│   │   │   └── val2017/
│   │   ├── lvis_train.pkl
│   │   └── lvis_val.pkl
│   ├── PACO-Part/
│   │   ├── coco/
│   │   │   ├── train2017/
│   │   │   └── val2017/
│   │   ├── paco/
│   │   │   ├── paco_part_train.pkl
│   │   │   └── paco_part_val.pkl
│   ├── Pascal-Part/
│   │   ├── VOCdevkit/
│   │   │   ├── VOC2010/
│   │   │   │   ├── JPEGImages/
│   │   │   │   ├── Annotations_Part_json_merged_part_classes/
│   │   │   │   └── all_obj_part_to_image.json

🥥 COCO-20ⁱ

Download COCO2014 train/val images and annotations:

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

Download train/val annotations: train2014.zip, val2014.zip.
Unzip and place both train2014/ and val2014/ under data/COCO2014/annotations/.

📸 FSS-1000

Download FSS-1000 images and annotations: FSS-1000.zip.

🦉 LVIS-92ⁱ

Download COCO2017 train/val images:

wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip

Download LVIS-92ⁱ extended mask annotations: lvis.zip. Unzip and place the .pkl files under data/LVIS/.

🧩 PACO-Part

Download COCO2017 train/val images (same as LVIS):

wget http://images.cocodataset.org/zips/train2017.zip
wget http://images.cocodataset.org/zips/val2017.zip

Download PACO-Part extended mask annotations: paco.zip. Unzip and place the .pkl files under data/PACO-Part/paco/.

🔩 Pascal-Part

Download VOC2010 train/val images:

wget http://roozbehm.info/pascal-parts/trainval.tar.gz
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2010/VOCtrainval_03-May-2010.tar

Download Pascal-Part extended mask annotations: pascal.zip. Unzip everything into data/Pascal-Part/VOCdevkit/VOC2010/.

Optional datasets

We do not use these datasets in SANSA experiments, but they are supported (datasets/deepglobe.py, datasets/isic.py, datasets/lung.py) and may be useful for others.

🌍 DeepGlobe

Download the dataset from the DeepGlobe. Expected layout:

SANSA/
├── data/
│   ├── Deepglobe/     
│   │   ├── image/                                  
│   │   └── filter_mask/

🩺 ISIC 2018

Download the dataset from the ISIC Challenge 2018. Expected layout:

SANSA/
├── data/
│   ├── ISIC/     
│   │   ├── ISIC2018_Task1_Training_GroundTruth/
│   │   ├── ISIC2018_Task1-2_Training_Input/

🫁 Chest X-ray

Download the dataset from the Chest X-ray. Expected layout:

SANSA/
├── data/
│   ├── LungSegmentation/ 
│   │   ├── CXR_png/
│   │   └── masks/