Ant Weber's Length Detection Using YOLOv11

This project presents a solution for detecting the Weber's length of ants using the YOLOv11-pose model. The workflow involves dataset preparation, model fine-tuning, and evaluation. The solution includes Python scripts and Jupyter Notebooks to facilitate model training, validation, and detection.

---

Project Overview

The Weber's length, a key morphological measurement of ants, can be extracted by identifying specific body landmarks and scaling the detected measurements. This project automates the detection and quantification process using YOLOv11 fine-tuning, making it efficient and reliable for biological research.

---

Project Structure

1. Dataset and Annotations

• annotations.csv: CSV files containing ground-truth bounding boxes and labels for the dataset.

2. Python Scripts

• create_dataset.py

  • Contains functions to prepare and preprocesses the dataset for YOLOv11 fine-tuning.
  • Handles annotation formatting and data splitting into training, validation, and test sets.

• utils.py

  • Provides utility functions used across the project, data visualization tools, scale bar detection function, etc ...

3. Configuration File

• hyp.yaml
  • Hyperparameter configuration file for YOLOv11 fine-tuning, specifying learning rates, epochs, etc ...

4. Jupyter Notebooks

• main_exp.ipynb

  • Contains the main training pipeline for the YOLOv11 model.
  • Fine-tunes the model using the prepared dataset and saves the best weights (best.pt).

• measurement.ipynb

  • Measures the results of the model with respect to ground-truth measurements.
  • Calculates the Weber's length based on detected landmarks.

• scale_bar_detection.ipynb

  • Detects and quantifies scale bars within input images to ensure proper measurement scaling using an image processing algorithm.

• text_detection.ipynb

  • Implements text detection (e.g., scale labels) from input images using easyOCR.

• DeepLSD_scale_bar_detection.ipynb

  • Implements scale bar detection using the DeepLSD project. (Make sure that the directory of DeepLSD is installed in the project folder and that the pre-trained DeepLSD weights, deeplsd_md.tar, are inside the project's folder). https://github.com/cvg/DeepLSD

---

Workflow

Go to step 3 if you want to test our model (.pt)

1. Scale Bar Detection:
   Run scale_bar_detection.ipynb (or using deepLSD) to detect and get the positions of the start and end of the scale bar.

2. Model Training:
   Run finetuning.ipynb to prepare the dataset and fine-tune the YOLOv11 model and save the best weights.

3. Evaluation:
   Use measurements.ipynb to compare the predicted Weber's length with ground-truth landmarks and assess the performance of the model.

4. Text Detection: If necessary, detect and get the textual annotation inside an image using text_detection.ipynb.

---

Requirements

• Python >= 3.8

• PyTorch >= 1.10

• ultralytics

• OpenCV

• easyocr

• NumPy

• Pandas

• Matplotlib

• Make sure thorax_only.csv and the image folder "ants" are inside the project's folder

• Add our model weights (.pt file) to the folder to test our finetuned model.

Install dependencies using: pip install -r requirements.txt

Reproducibility

We provide a drive with checkpoints for all our trained YOLO models, as well as our dataset splits (train\val\test) https://drive.google.com/drive/folders/1ElyO_maC5C08JpuWhXqTZu9ZBB9VzCWa?usp=sharing