Probe detection
A YOLOv8 nano detector for inspection probes, built as a technical assignment for Flyability.
Details
- Period
- 2024
- Status
- Completed
- Code
- Private repository
- Stack
- YOLOv8, Python
Case study
Task
Detect probes in images, fast enough for near real-time use.
Approach
A single-class YOLOv8 nano model, chosen for its small size and speed, fine-tuned on annotated images with bounding boxes in COCO format. The data was split 64/16/20 into training, validation and test sets. Training used AdamW at a learning rate of 0.002, batch size 16, for 50 epochs.
Results
| Precision | Recall | mAP50 | mAP50-95 | Inference |
|---|---|---|---|---|
| 0.995 | 1.0 | 0.995 | 0.888 | 58 ms per image |
The dataset was simple and consistent, so the model converged quickly. Next steps would be data augmentation for robustness, and quantization and pruning to run on edge devices such as NVIDIA Jetson boards.