Animal Detection & Tracking with YOLO

This video shows how to implement YOLOv8 in a project where you don't need pose-estimation data. Many times in neuroscience experiments, pose-estimation data is unnecessary (e.g., open-field activity tracking, total time spent in an ROI, etc). For such applications, implementing an efficient model, like YOLOv8, could save you a lot of time. In this tutorial, I provide you a step-by-step guide on how to extract frames from your videos, annotate them for YOLO model training, and then use the trained model on new open-field mouse videos.

Since YOLOv8 is a versatile model, it can easily be used for other object segmentation and classification needs. Its applications already exist in microscopy (cell classification, segmentation, etc.).

# Code snippets used in this video:

1. Create Anaconda Environment
`conda create -n yolov8 python=3.9`

2. Extract Frames from Behavior Videos
` ffmpeg -i input.mp4 -vf "fps=1/2" -frames:v 40 output_%02d.png `

3. Create a New Folder (label: "images") and Move Extracted Frames into the "images" folder
3.1. Create another folder and label it "labels"

4. Install Image Annotation Tool
` pip install labelImg `

5. Run Annotation Tool
`labelImg `
# Change the Save Directory
# Make sure that the source is selected to be YOLO and not Pascalvoc

6. Annotate your Object by creating a rectangle and labeling it.

7. Install ultralytics
` pip install ultralytics `

8. Upgrade/Install Pytorch
# Copy code from here and install it into your environment:
` https://pytorch.org/get-started/locally/ `

# Example: ` pip3 install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 `

9. Create "val" and "train" folders

10. Subdivide "val" folder by making two more folders (i.e., "images" and "labels" folders, respectively; these will be used for evaluation of model after training).
10.1. Move some of your training images from the training folder into the "images" folder under the val directory. Move the corresponding labels from the training folder into the "labels" folder under the val directory. [delete the images and labels that you are moving from the training directory to the val subfolders so that the algorithm won't see these images or labels while training]

11. Create "data_custom.yaml" using text editor [copy the code below into "data_custom.yaml" file using text editor]

train: "Path to your train images folder"
val: "Path to your val folder"
nc: 1
names: ["Mouse"]

12. Run YOLO for training on your images
` yolo task=detect mode=train epochs=500 data=data_custom.yaml model=yolov8m.pt imgsz=640 batch=8 `
# best-trained model will save in a folder called "weights"

13. Run YOLO Custom Model on New Videos with Labeled Predictions
` yolo task=detect mode=predict model=best.pt show=True conf=0.5 save_txt=True source= "video path/name" `

14. Run YOLO Custom Model on New Videos without Labels and Confidences Printed on the Video
` yolo task=detect mode=predict model=best.pt show=True conf=0.5 save_txt=True source= "video path/name" hide_labels=True hide_conf=True" `

Ultralytics Github: https://github.com/ultralytics/ultralytics

Видео Animal Detection & Tracking with YOLO канала NeuroTech