Skip to content

Latest commit

 

History

History
46 lines (38 loc) · 1.99 KB

DATA.md

File metadata and controls

46 lines (38 loc) · 1.99 KB
Raw

Data Preparation

Preparation for C20K dataset

  • Download the COCO 2014 dataset and put it in data/.
  • Download mixed_lite.mat and mixed_image_paths.mat for C20K dataset here and put it in data/coco_train_20k/mixed then run the following code in Matlab from the LOD folder to modify the image paths according to your LOD path.
setup;
imdb = load(fullfile(LOD_ROOT, 'data/coco_train_20k/mixed/mixed_image_paths_original.mat'));
imdb.image_paths = cellfun(@(el) fullfile(LOD_ROOT, 'data', el), imdb.image_paths, 'Uni', false);
savefile(fullfile(LOD_ROOT, 'data/coco_train_20k/mixed/mixed_image_paths.mat'), imdb);

Use your own dataset

LOD requires the data to be organized as follow:

data/
|   dataset1/
|   |   class1/
|   |   |   class1_lite.mat
|   |   |   class1_image_paths.mat
|   |   class2/
|   |   |   class2_lite.mat
|   |   |   class2_image_paths.mat
|   |    ...
|   dataset2/
|    ...

Through the project, we suppose that all datasets have a single class mixed. Its information is contained in two .mat files, mixed_lite.mat and mixed_image_paths.mat. mixed_lite.mat is a struct with the following fields:

bboxes: (n x 1) cell, each cell is a (K x 4) matrix containing ground-truth bounding boxes of the images in the dataset.
        The bounding boxes are in format [x1,y1,x2,y2], inclusive.
        THIS FIELD IS ONLY USED FOR EVALUATION. YOU DO NOT NEED IT TO RUN THE METHOD.
        IN THE CASE GROUND TRUTH IS NOT AVAILABLE, YOU CAN SET bboxes=cell(n,1)
images_size: (n x 1) cell, containing images' size (height x width).

and mixed_image_paths.mat is a struct with a single field

image_paths: (n x 1) cell, containing paths to the images in the dataset.

To use your own dataset, create a mixed_lite.mat and a mixed_image_paths.mat with the fields above and put it in the right place. See the .mat files in data/coco_train_20k/mixed for an example.