DOTA Aerial Object Detection - YOLOv11 Demo Project

By: Javier H.

Colab link: https://colab.research.google.com/drive/1mb9WWhO9-k271FvuFEUB8LfRGbXbcCN-?usp=sharing

---

This project demonstrates how to preprocess and train the YOLOv11 model on aerial imagery using sample data from the DOTA v1.5 dataset.

Dataset

• DOTA (Dataset for Object deTection in Aerial images): A large-scale benchmark for detecting oriented objects in aerial images.
• Version: v1.5 (used for demonstration; v2.0 is not in Google Drive); Official site
• Note: YOLOv11 is typically trained on DOTA v1.0, but this project focuses more on data transformation and preprocessing, not benchmarking accuracy.

Model

• Model: YOLOv11 Nano variant
• Input Size: 640×640 (default)

Preprocessing & Data Transformation (Main Focus)

This project shows one of the processes I take to convert datasets into YOLO-compatible format, especially for datasets similar DOTA with oriented bounding boxes (OBB) or varying image resolution and sizes.

• Tiling: Images are split into overlapping tiles of 640×640 pixels (with black padding as needed)
• Label Adjustment:
  • Translates OBB annotations to fit each tile
  • Clamps bounding boxes to tile edges
  • Converts polygon OBBs to YOLO-oriented format (normalized coordinates)
  • Converts class names to integer indices
• Why Tiling?:
  • Prevents distortion from resizing non-square images
  • Preserves detail in high-resolution images
  • Improves detection of small, dense aerial objects

Training

• Training is done using the transformed dataset
• Evaluation metrics, training logs, and visual results (e.g., predictions, confusion matrix, loss curves) are produced to monitor performance

This project focuses on the process and pipeline in training the YOLOv11 model.

Experimentation

Experimentation is crucial in machine learning, so it is natural that there is trial and error present in the code. This is a Jupyter notebook after all. As a result, some cells have duplicate import statements due to running them individually for debugging purposes.

Files

• The model_training_[date].txt with the latest date has the latest model training runs on a T4 GPU (with a short accompanying video during training).
• The JavierH_YOLOv11_DOTA_v1_5.ipynb has the Jupyter Notebook with the shared Google Colab link attached.
• The yolo_runs/obb_exp/ folder contains the data generated during the latest run of YOLOv11.