The YOLOv1 algorithm implemented in Tensorflow 2.4 and Keras 2.6, with a network architecture consisting of a pretrained network and a fully connected head from Keras applications architectures. Image augmentation was not used in this project and for good validation/test results it is needed to use image augmentation. In the future, image augmentation will be implemented.
Weights trained on ResNet50 with the YoloLoss loss from losses.py: https://drive.google.com/file/d/1NwRcqHB_dGR5gqqtt_YHrHgSwbqGOObj/view?usp=sharing The hdf5 file should be placed in output folder.
This file contains the VOCdataset class with the load_dataset() function. This function returns the boxes, labels, number of objects, image_path for each image. The dataset directory should contain a directory with images and a directory only with xml annotations. Each annotation is read independently and there are no csv files used.
Contains multiple functions. First the transform() function that transforms the information for each image into a yolo format array of size 30,7,7. If you want to train the algorithm on a different dataset with different classes, the values for the variables must be updated. (Mostly where there is a number with 20, should be replaced with the new number of classes)
Choose to visualize prediction with transform_from_grid.py without nms, or with trasnform_with_nms. Model can be loaded through load_model function, preprocessing for predictions should be adapted to the trained model preprocessing.
If input size is changed from (224,224), the W and H in yolo_loss should also be changed. When changing network, also change preprocessing. When changing the number of grids, also change the number of grids for yolo_loss, respectively S.
Update W and H when using different input size than (224,224), also update S when using other number of grids than 7.
This application was created as a bachelor thesis project and as an education tool. Even though the code is thought for 20 classes, with minor tweekings it can be adapted for new datasets. There are only 2 bounding box predictions per cell, but this can also be change by adapting the B variable and also changing some parts in the transforms file and the loss files.