Yolo Exploration

Get started

1. Clone yolov7 repo git clone https://github.com/WongKinYiu/yolov7.git
2. Download yolov7 trained weights from here and put it in yolov7 directory. wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt
3. Install all necessary dependency using pip install requirements.txt. You need Cuda to install talk to your GPU.
4. Put training_pipeline.ipynb, custom_scripts, model_prediction_custom.ipynb, visulization_pipeline.ipynb in yolov7 directory

Train the model

1. training_pipeline.ipynb contains data_preprocess and model training module.
2. Input data:

• tif file: a RGB image areal with 5cm resolution.
• geojson bounding boxes in the same projection as the tif

3. The data will preprocess in first phase and stored images and their label in txt format on given path.
4. After the data preprocess and if the aug flag is TRUE then data augmentation will start. After the augmentation the data will be divided in training and validation sets. If aug flag is FALSE then data will directly be split into the two sets (ratio is training = 80, validation = 20).
5. After the creation of the validation_set and training_set model the training will start. You can change the parameter in yolo class for training.
6. All paramters (epochs) can be changed in config.py.
7. In the weights directory best.pt is best model. But you can save it on different location.

Notes

• data_augmentation.py file is required for data augmentation.
• default cropping_size = 2500. This means the image will be croped from main tif file with height, width = 2500. Set cropping size according to your needs. This is required if the input data (resolution) changes.
• For a better model accuracy keep crop_size same for model training as well as model prediction.

Prediction

1. prediction_with_geojson.ipynb will accept (tif file and model_weights) and save predicted images in given directory. The model .pt file is required for the model prediction and the path for model weights needs updating to the location of best.pt.
2. The prediction process script will generate crops from the tif file and predict and store the output in the destination folder. If the object is detected a bounding box is drawn on the image.
3. The script will create a geojson file with the realworld coordinates of the the detected objects.

Visualization

1. visualization_pipeline.ipynb will accept tif file and geojson file as input and visualize bounding box. You can use this script to verify input/predicted data.

Party !!!