Repository files navigation

Working on README. Please check later!!!

Here is the section you can follow to create your database, it is the way we have done it but it isn't the only way.

We took the images with the drone with the scriptsdataset_creator.py and by pressing the key "i", you just had to change the folder where you want it to save images. Note : the number max of images are 999 999, for the entire dataset creation. If you want to extend it you must change theindex() methode.As background images have been added at the end, the current function have to be updated for the images without object when creating .txt, thus you'll want to put the background images at the end of your pool.

The label format is a .txt file where each lines match one object.The format of a line is :class x_center y_center width height , where coordinates must be normalized by the dimension of the image and calss numbers are zero-indexed.

Example : 0 0.48225308641975306 0.35133744855967075 0.21682098765432092 0.34053497942386834

We search the easiest way to annote images with bounding boxes. We choose to use LabelMe from MIT. There is a repo githubhere to install it with pip.This annotate tool output a JSON format for the labeling, so we have a function callcreateLabel() inutils_dataset.py to allow you do to the conversion to YOLO format.As said earlier we need to adapt the function to the abcense of object during the annotation, so for now you need to add empty .txt at the end to match bg images.

We double our database with an offline data augmentation, which is a simple horizontal flip, to do so use the funtioncreate_augmented_data().

We want this arborecence :

For the training we need to split our pool in two (or three) part, to do so you can use thesplit() function that split 90/10 for training and validation.

Watchthis video to see our results.