One eye. All the depth you need.
Real-time monocular depth estimation on a Raspberry Pi 5, built in 48 hours at the European Defense Tech Hackathon Porto, 2026. 2nd place.
📊 View the pitch deck →     Slide presentation, deployed via GitHub Pages.
🛠️ See the engineering repo →     Training code, ablation experiments, ONNX export, and Pi 5 inference.
A drone needs to know how far away things are. The conventional answers all carry a tax: stereo cameras need calibration and baseline, time-of-flight sensors are expensive, LiDAR adds weight and cost. Cyclops Vision asks whether a single ordinary camera plus a small neural network can deliver “good enough” depth for situational awareness on a power-and-weight-constrained platform — and at what cost.
The result: a fine-tuned RTMonoDepth_s model (~1.3 M parameters), distilled from Depth Anything V3 Nested as the teacher, running as a ~5 MB ONNX model at 5–10 FPS on a Raspberry Pi 5 with no PyTorch runtime dependency.
| Component | Cost |
|---|---|
| Compute (STM32MP25 + memory) | 25 € |
| Camera (Sony IMX296 + optics) | 60 € |
| Power & I/O | 20 € |
| PCB & assembly | 18 € |
| Housing & mechanical | 25 € |
| Calibration & QC | 27 € |
| Total BOM (at 1,000 units) | 175 € |
Scales to ~105 € at 10k units.
This repo is the pitch artifact: the slide deck, branding assets, and presentation site. The actual training and deployment code lives in the companion engineering repo.
.
├── Presentation/ # HTML slide deck (deployed via GitHub Pages)
├── logo/ # Project logo (HTML / assets)
└── animation/ # Supporting animation assets
Built in 48 hours by Niclas, Luca, and Lars Husemann at the European Defense Tech Hackathon, Porto, 2026.