How to install datasets

October 17, 2022 · View on GitHub

To start with, you can download the following datasets and store them under $DATA directory. In our case we use data/ as the default and if you want to use a different path, just make sure you define it in the arguments of the python scripts that store or load from there.

The file structure looks like:

data/
|–– caltech-101/
|–– eurosat/
|–– cct20/

Datasets list:

The instructions to prepare each dataset are detailed below. To ensure reproducibility and fair comparison for future work, we provide fixed train/val/test splits. The fixed splits are either from the original datasets (if available) or created by the authors of CoOp and ourselves (last 6 datasets in the list).

Standard Datasets

Caltech101

The directory structure should look like

caltech-101/
|–– 101_ObjectCategories/
|–– split_zhou_Caltech101.json

OxfordPets

The directory structure should look like

oxford_pets/
|–– images/
|–– annotations/
|–– split_zhou_OxfordPets.json

StanfordCars

The directory structure should look like

stanford_cars/
|–– cars_test\
|–– cars_test_annos_withlabels.mat
|–– cars_train\
|–– devkit\
|–– split_zhou_StanfordCars.json

Flowers102

The directory structure should look like

oxford_flowers/
|–– cat_to_name.json
|–– imagelabels.mat
|–– jpg/
|–– split_zhou_OxfordFlowers.json

Food101

The directory structure should look like

food-101/
|–– images/
|–– license_agreement.txt
|–– meta/
|–– README.txt
|–– split_zhou_Food101.json

FGVCAircraft

The directory structure should look like

fgvc_aircraft/
|–– images/
|–– ... # a bunch of .txt files

SUN397

The directory structure should look like

sun397/
|–– SUN397/
|–– split_zhou_SUN397.json
|–– ... # a bunch of .txt files

DTD

The directory structure should look like

dtd/
|–– images/
|–– imdb/
|–– labels/
|–– split_zhou_DescribableTextures.json

EuroSAT

The directory structure should look like

eurosat/
|–– 2750/
|–– split_zhou_EuroSAT.json

UCF101

  • Create a folder named ucf101/ under $DATA.
  • Download the zip file UCF-101-midframes.zip from here and extract it to $DATA/ucf101/. This zip file contains the extracted middle video frames.
  • Download split_zhou_UCF101.json from this link.

The directory structure should look like

ucf101/
|–– UCF-101-midframes/
|–– split_zhou_UCF101.json


Challenging Datasets

The following datasets make up the challenging datasets used in the SVL-Adapter paper. We provide the splits we used in a .json file for the shake of benchmarking and comparison. When available, the train/test splits follow the ones provided by the original curators of each dataset.

FMoW

The Functional Map of the World (FMoW) dataset presented in this paper contains thousands of satellite images which are labeled based on the functional purpose of the building or land they contain. We use the fMoW-rgb version of the dataset and keep a subset of the data (defined in split_FMOW.json) for efficiency.

  • Create a folder named fmow/ under $DATA.
  • Download the images along with bounding box and annotations https://github.com/fMoW/dataset for both train/val and test subsets.
  • Extract these files under $DATA/fmow/.
  • Download split_FMOW.json from this link.

The directory structure should look like:

fmow/
|–– train/
|–– test/
|–– split_FMOW.json

OCT

This dataset contains thousands of validated Optical Coherence Tomography (OCT) described and analyzed in this paper. The images are split into a training set and a testing set of independent patients with images having labels from 1 of the 4 following categories: CNV, DME, DRUSEN, and NORMAL.

  • Create a folder named oct/ under $DATA.
  • Download the images from https://data.mendeley.com/datasets/rscbjbr9sj/3 (the image labels are indicated by the name of the folder they are in).
  • Extract these files under $DATA/oct/.
  • Download split_OCT.json from this link.

The directory structure should look like:

oct/
|–– train/
|–– test/
|–– split_OCT.json

Camera Trap datasets

A large repository of camera trap data can be found at lila.science, including Caltech Camera Traps (CCT20), Island Conservation Camera Traps (ICCT) and Snapshot Serengeti datasets which were used for to evaluate SVL-Adapter across challenging task. For each camera trap dataset examined, we extract the bounding boxes around the object of interest given their availability. Note: If bounding box annotations are not available for a camera trap dataset, regions around animals can be extracted pretty accurately by utilizing the [MegaDetector].

CCT20

The directory structure should look like:

cct20/
|–– train_images/
|–– cis_val_images/
|–– cis_test_images/
|–– trans_val_images/
|–– trans_test_images/
|–– split_CCT20.json

ICCT

The directory structure should look like:

icct/
|–– train_images/
|–– cis_val_images/
|–– cis_test_images/
|–– trans_val_images/
|–– trans_test_images/
|–– split_ICCT.json

Serengeti

Note: We use a subset of this dataset (defined in split_SERENGETI.json).

  • Create a folder named serengeti/ under $DATA.
  • Download the images that have bounding box information and annotations available https://lila.science/datasets/snapshot-serengeti.
  • Extract these files under $DATA/serengeti/.
  • Download split_SERENGETI.json from this link.

The directory structure should look like:

serengeti/
|–– train/
|–– test/
|–– split_SERENGETI.json
|––

MMCT

  • This dataset is collected in the Maasai Mara region in Kenya for the Biome Health Project which is funded by WWF UK. The dataset not public yet. We will add a link to the data and the splits used as soon as it becomes available.