YOLO uses convolutional neural networks instead of the region-based methods employed by alogirithms like R-CNN. The convolutional network Only Look Once, ie it requires only one forward pass through the neural network to make predictions. It makes two predictions, the class of object and the position of the objects present in the image. YOLO divides an image into nxn regions and predicts whether a target class is present in each region or not. It also predicts the bounding box coordinates of the target class. Non-max suppression is used to prevent same object being detected multiple times. Refer this video for a better, in-depth understanding of YOLO models. The original YOLO paper could be accessed here and YOLOv5 repo could be found here
YOLOv5 is available over this repo.
The data in CycleGAN format is available here as a zip file.
We use it to train the model to work with the Signature detection use case.
YOLOv5 requires the dataset to be organized in a particular structure.
dataset
|-> images
|-> train (.jpg files)
|-> valid (.jpg files)
|-> labels
|-> train (.txt files)
|-> valid (.txt files)
Use this notebook to train and test the YOLOv5 model.
Clone the official [YOLOv5 repo] and install the requirements using the requirements.txt file.
We need to create a tobacco_data.yaml and add the path of training train: and validation valid: directories, number of classes nc: and class names ['DLLogo', 'DLSignature'] and add this file to the yolov5 directory we cloned.
Now, we have to select a pre-trained model from the available pre-trained checkpoints. These model.yaml files are present inside yolov5\models. I have used yolov5x for performing my experiments.
Training arguments
--img 640 is the width of the images.
--batch - batch size.
--epochs - no of epochs
--data - Dataset.yaml (tobacco_data.yaml) path
--cfg models/model.yaml is used to set the model we want to train on. I have used yolov5x.yaml, more information could be found here.
To Train the model, run the following line.
!python yolov5/train.py --img 640 --batch 8 --epochs 10 --data tobacco_data.yaml --cfg models/yolov5x.yaml --name Tobacco-run
A maximum of 10 epcoh was tested, time lacking
Each line of the txt file contains class: x_center, y_center, width, height.
Box coordinates are in normalized form. x_center and width are divided by image_width and y_center and height are divided by image_height to get a value between 0 to 1.
Given for the project:
- Test
- Train
- Train-xml
Produced:
Scaled: originals images are resizd to reduced training time.
tobacco_yolo_format: scaled images are preprocess and placed in correct directories to be used for yolov5
First, use 'preprocess_to_yolo' to prepare your data, just adapt pathes. Could output 'tobacco cleaned.csv' Then you can use 'launch yolo'. Google colab could be useful
check my google colab to see how far I went: model trained with 10 epoch and tested on test files provided https://colab.research.google.com/drive/1PIg7aLsZ8439Inz49lPfm8StOwaKPiAP?usp=sharing
thanks to https://github.com/amaljoseph/Signature-Verification_System_using_YOLOv5-and-CycleGAN