SCV_Project

In order to use / evaluate our work, a clear description of our contribution / code is provided below.

Demo

One can use the images in the directory demo for a short demo. The images includes a variety of images, including plant images the network has not trained on.

python ./yolov3/detect.py --weights ./yolov3/runs/train/exp/trained_weights.pt --source ./leaf_data/demo_images --img-size 416

The defaults have been pre-configured along with a path to two example image files.

Setup data

In the create_dataset.py file, we call the following method to prepare the data for training. 0. Create a directory called leaf_data, which includes two directories images and labels, which again both include three directories train, validate, and test.

1. Call moveImagesFromOriginalDataset() to copy the dataset images into the network training directory.
2. Call createAugmentedDataset(num_train, num_test, num_val) in order to create the augmented images. Created by Yaniv
3. Call moveAugmented() to then move the augmented label-images (the images based on which we create the bounding boxes) to the label images directory. Created by Yaniv & Sterre
4. Call setupCreateLabels() to create the bounding boxes txt files, for all un-augmented images, in the right directory (along the label-images). Created by Yaniv & Sterre
5. Call setupCreateLabelsAugmented() to create the bounding boxes txt files, for all augmented images, in the right directory (along the label-images). Created by Yaniv & Sterre
6. Call prepareTxts() to create a train.txt, test.txt and val.txt, which contain a list of the files to be used in the network training for train / test / validation. Created by Yaniv & Sterre
7. Call prepareTxtsAugment() to do the same for the augmented images. Created by Yaniv & Sterre

Test bounding boxes:

In the file show_bounding_boxes.py we have implemented functionality to show bounding boxes, for bounding box testing purposes. Created by Sterre

Train

In order to train, we have imported from https://github.com/ultralytics/yolov3.git, created a dataset yaml file (yolov3/data/leaf.yaml) which specifies the dataset information (number of classees, class names, source for train / test / validation).

To train the network, we call:

python train.py --weights yolov3-tiny.pt --data data/leaf.yaml --img-size 416 --epochs 60

In order to avoid a large number of passed parameters, all of the above have been changed into the parsing-default, so one can also only call: python train.py

The results, along with the trained weights, are automatically generated into runs/train/expX.

Detection

In order to detect, we call:

python ./yolov3/detect.py --weights path_to_weights --source images_to_detect_on --img-size 416

The trained weights are available in the path "./yolov3/runs/train/exp/trained_weights.pt"

Additional files

YOLOv3_tutorial_based.py and util.py are a yolo-architecture and detector implementation based on https://blog.paperspace.com/how-to-implement-a-yolo-object-detector-in-pytorch