Part 2: NASA Crater Detection
This is Part 2 of a two-part series on crater detection using deep learning. Missed Part 1? Catch up here →
🚀 Overview
In Part 1, we built a YOLOv10 + Ellipse R-CNN pipeline to detect and localize lunar craters.
This post focuses on:
- 🌀 Rim-fitting using Ellipse R-CNN
- 📊 Evaluation metrics and precision challenges
- ⚙️ System integration and resource constraints
- 🔍 Final takeaways and next steps
📺 Watch a 1-minute summary of the project:
🧠 From Boxes to Rims: LunarLens Pipeline
Our two-stage detection system — LunarLens — works like this:
- YOLOv10 detects crater bounding boxes quickly.
- Ellipse R-CNN then fits an ellipse to the rim of each crater.
This hybrid system balances speed with shape accuracy, ideal for autonomous space navigation.
🌀 Postprocessing with Ellipse R-CNN
Bounding boxes from YOLOv10 are cropped and passed to Ellipse R-CNN, which predicts:
- (x, y) center
- Major & minor axes
- Orientation
This gives us a mathematically accurate rim outline.
Crater 1: (100.3, 40.8), (101.5, 39.6), ...
Crater 2: (463.2, 80.9), (462.3, 81.4), ...
📊 Evaluation & Metrics
We assessed detection quality at IoU thresholds of 50% and 70%, using AP, precision, and recall.
| Metric | IoU = 50 | IoU = 70 |
|---|---|---|
| AP | 0.111 | 0.003 |
| Recall | 0.145 | 0.010 |
| Precision | 0.337 | 0.042 |
🧠 Key takeaway: Precision drops sharply at high IoU — highlighting how tight rim localization is the hard part, not detection.
🧪 System Performance Under Constraints
We were tasked with running this system on CPU-only hardware with Raspberry Pi–level specs.
| Constraint | Result |
|---|---|
| Image Size | 5 MP |
| Runtime | ~20 sec / image |
| RAM | < 4 GB |
| Hardware | No GPU |
🔍 Final Thoughts & Next Steps
- ✅ YOLOv10 was great for rapid detection, but rim precision required postprocessing.
- ✅ Ellipse R-CNN helped convert detections into usable geometric inputs.
- ⚠️ Evaluation at high IoU exposed the need for better shape-fit training.
Next steps:
- Fine-tune Ellipse R-CNN on lunar tile crops
- Apply weighted evaluation based on crater size
- Improve YOLO thresholding and ensemble strategies
🤝 Acknowledgments
Built by: Anekha Sokhal, Tian Le, Henry Tran, Juan Hevia, Madeleine Harrell, Jeremy Xu Mentored by: Kyle Smith (NASA), Arko Barman, Ananya Muguli
The team at NASA Johnson Space Center, Houston
💬 Have questions, or want to explore vision systems for autonomous spacecraft?
Feel free to reach out — I’d love to chat!