Training YOLOv3 on a Custom Dataset

This guide provides a step-by-step workflow for training YOLOv3 to detect a custom object, in this case, a specific cow. The workflow combines both local and cloud-based computational resources for optimal performance.

Kaggle Dataset

The full dataset, including images, YOLO labels, and Darknet configuration files, is available on Kaggle.

---

Table of Contents

• Introduction
• Hardware and Software Environment
  • Workflow Overview
• Preprocessing Input Data
• Labeling Input Images
  • Manual Labeling with LabelImg
  • Semi-Automated Labeling using Object Tracking
• Configuring YOLOv3
  • Download YOLOv3
  • Modify YOLO Configuration
• Training YOLOv3 on Google Colab
• References

---

Introduction

YOLO (You Only Look Once) is a state-of-the-art object detection model capable of identifying and localizing objects in images or video frames. Pre-trained versions are available for common objects (COCO dataset, 80 classes). In this project, we train YOLOv3 to detect a specific cow, but the workflow can be adapted for any object.

We use:

• Darknet: Open-source neural network framework for YOLO.
• Google Colab: Linux-based GPU environment for training.
• Local machine: Windows-based environment for preprocessing, manual labeling, and tracking using OpenCV, LabelImg, and dlib.

---

Hardware and Software Environment

Workflow Overview

1. Image and Video Preprocessing

   • Done locally using OpenCV and FFmpeg.
   • Includes resizing, frame extraction, format conversion, etc.

2. Manual or Semi-Automated Labeling

   • Use LabelImg for manual labeling.
   • Semi-automated labeling: Track objects in video with dlib, save bounding boxes as YOLO labels.

3. Training and Testing YOLOv3

   • Conducted on Google Colab with GPU support.
   • Custom anchor sizes calculated and YOLOv3 configured for your dataset.

4. Interoperability

   • Data is exchanged between local machine and Google Drive to manage preprocessing, training, and post-processing.

---

Preprocessing Input Data

YOLO requires both images and corresponding .txt label files. Each line in the label file contains:

<class_id> <center_x> <center_y> <width> <height>

• <class_id>: Object class (0 if single class)
• <center_x> & <center_y>: Center of bounding box relative to image dimensions
• <width> & <height>: Width and height relative to image dimensions

---

Labeling Input Images

Manual Labeling with LabelImg

Steps:

1. Take snapshots of the target cow from video using FFmpeg:

ffmpeg -i input_video.mp4 -vf "fps=0.5,scale=1920:-1" -q:v 1 snapshot%d.png

2. Label snapshots in LabelImg:

# Navigate to LabelImg folder
cd C:\your_path
# Run LabelImg
python labelImg.py

• Draw bounding boxes and assign class.
• Save annotations (.txt) for YOLO.
• If LabelImg assigns a different class ID (e.g., 15), adjust it to 0 manually or via provided script.

Semi-Automated Labeling using Object Tracking

1. Select a clear frame of the target object and manually annotate the bounding box.
2. Use dlib correlation tracker to track the object in video.
3. Save frames and bounding boxes at intervals, converting coordinates from dlib format to YOLO format.
4. Save snapshots and .txt labels with consistent filenames.

---

Configuring YOLOv3

Download YOLOv3

Download Darknet YOLOv3 from AlexeyAB/darknet.

Modify YOLO Configuration

• .cfg file:

  • batch = 64
  • subdivisions = 64
  • max_batches = 2000 (for 1 class)
  • classes = 1
  • filters = (classes + 5) * 3 = 18
  • Update [yolo] layers with calculated anchors.

• Custom anchors (from k-means clustering):

20 27, 16 23, 16 31, 18 27, 15 21, 19 26, 15 27, 15 25, 13 23

• obj.data:

classes = 1
train = /content/darknet/train.txt
valid = /content/darknet/valid.txt
names = /content/darknet/obj.names
backup = /content/darknet/backup/

• obj.names:

cow13

• train.txt / valid.txt:

/content/darknet/data/train/image1.jpg
/content/darknet/data/train/image2.jpg
...

---

Training YOLOv3 on Google Colab

Upload the following files to Google Drive:

• backup folder
• obj.data
• obj.names
• train.txt
• valid.txt

Run in Colab:

# Add execute permission
!chmod +x /content/drive/MyDrive/darknet/darknet

# Change directory
%cd /content/drive/MyDrive/darknet

# Start training
!./darknet detector train /content/drive/MyDrive/darknet/data/obj.data /content/drive/MyDrive/darknet/cfg/yolov3.cfg -dont_show

• Save trained weights to backup folder.
• Test detection on a video:

!./darknet detector demo /content/drive/MyDrive/darknet/data/obj.data \
/content/drive/MyDrive/darknet/cfg/yolov3.cfg \
/content/drive/MyDrive/darknet/backup/yolov3_best.weights \
-dont_show -out /content/drive/MyDrive/outputVideo.mp4 \
/content/drive/MyDrive/cutVideo.mp4

• Detect on a single image:

!./darknet detector test /content/drive/MyDrive/darknet/data/obj.data \
/content/drive/MyDrive/darknet/cfg/yolov3.cfg \
/content/drive/MyDrive/darknet/backup/yolov3_best.weights \
/content/drive/MyDrive/cow3.png -dont_show -out /content/drive/MyDrive/outputImage.jpg

---

References

1. Redmon, J., Divvala, S., Girshick, R., Farhadi, A., You Only Look Once: Unified, Real-Time Object Detection, 2016.
2. Lin, T.Y., et al., Microsoft COCO: Common Objects in Context, 2014.
3. FFmpeg
4. LabelImg
5. AlexeyAB Darknet
6. Kaggle dataset for YOLOv3 Cow Detection: https://www.kaggle.com/datasets/gayebartos/yolov3-cow-detection