This repository contains scripts and results for using YOLOv3 object detection model to detect tips and open collars in maize plants. Tips and open collars can inform scientist and farmers about the age and health of the plant. It can also aid in extracting other useful information from the images such as leaf length and internode length.
First, follow the instructions to clone the YOLOv3 repository which also includes pretrained models. We used a pretrained backbone and fine tuned the model on the maize dataset.
The dataset used to train and test the model can be found here.
We utilized google collab for this project so that we could keep the data on the cloud. The prepare_data collab notebook runs the user through processing the data. Update paths to match directory structure. The notebook run_yolo enables interactive training and testing of the model as well as generating labeled images.
Here are some of the output photos we got during the test run:
Clearly the model does a good job of finding most keypoints, however you can see that there are quite a few false detections. This is reflected in our test metrics curves below.
Training seems to be going well:
Testing gave decent results. However, as we could see in the output images, the recall is high but precision is lower resulting in a lower total mAP and F1 score:
