Add the required input paths to the main.py code and chaange the output paths if required. The code was developed using YOLO style annotation data and expects input annotations in the format <class name> <x> <y> <width> <height>, just like any YOLO architecture. The classes variable in the main.py can be changed to whatever classes are there in the dataset, the current list is according to the Pascal VOC dataset.
Input images:
| Image 1 | Image 2 | Image 3 | Image 4 |
|---|---|---|---|
 |
 |
 |
 |
Output mosaic augmented image with YOLO compatible annotations:
| Output Mosaic Image | Output with BBoxes |
|---|---|
 |
 |
The dataset.py function converts the given image and annotation directories into lists of images and their corresponding annotations. After which the main.py randomly samples 4 images from the dataset and passes it to the mosaic.py function which scales the four images and arranges them to fit in the desired output size and also changes the annotaion values accordingly.
Heavily inspired from this code which did not support YOLO style annotations