Lunar Terrain Object Detection Using the POLAR SIM Dataset

Abstract

This project presents a robust object detection system built to identify rocks and shadows on lunar terrains using the POLAR SIM dataset. The objective is to support autonomous navigation and scientific exploration on extraterrestrial surfaces by enabling real-time detection under challenging and diverse illumination conditions.

Methodology

1. Dataset Exploration & Annotation

   • Utilized the POLAR SIM dataset, a synthetic lunar imagery dataset featuring diverse terrain conditions.
   • Labeled classes: rock, shadow.

2. Model Selection

   • Chose two detection models for experimentation:
     • YOLOv5 Nano – one-stage, lightweight model for real-time use.
     • Faster R-CNN with MobileNet – two-stage detector for high-precision results.

3. Data Augmentation

   • Applied advanced augmentation techniques to increase robustness to lighting and orientation variability:
     • Rotation, Scaling, Brightness/Contrast Adjustments, Flipping.

4. Training Pipeline

   • Developed end-to-end pipelines using PyTorch and Albumentations for preprocessing.
   • Fine-tuned YOLOv5 Nano and Faster R-CNN models on the POLAR SIM dataset to enhance detection accuracy for rocks and shadows in diverse lunar terrain conditions.

5. Evaluation

   • Measured model performance using mAP (mean Average Precision) and class-wise accuracy.
   • Benchmarked both models on the same test set to compare performance under real-world constraints.

Key Contributions

• Developed a Dual-Model Object Detection System for classifying rocks and shadows on synthetic lunar terrain images.
• Achieved 91% accuracy with YOLOv5 Nano and 84% with Faster R-CNN, balancing speed and precision.
• Implemented Realistic Data Augmentation to simulate varied lunar surface conditions and improve generalization.
• Optimized for Edge Inference, with YOLOv5 Nano ideal for deployment on lightweight hardware like lunar rovers.
• Open-source Codebase structured for reproducibility and extensibility, suitable for research and production use.

Model Performance

Tech Stack

• Programming Language: Python
• Deep Learning Framework: PyTorch
• Detection Models: YOLOv5 Nano, Faster R-CNN (MobileNet)
• Computer Vision Tools: OpenCV
• Data Augmentation Library: Albumentations
• Dataset: POLAR SIM (Synthetic lunar terrain dataset)

Repository

🔗 GitHub Repository – SamarthKukreja/object_detection