GPU-accelerated desktop workspace for monocular depth estimation, interactive 3D visualization, semantic inspection, and 3D reconstruction.
Live RGB-to-depth inference with interactive depth-map and 3D point-cloud visualization.
Segmentation-assisted object inspection with object-level point extraction and 3D geometry.
- Live depth inference from images, video, USB cameras, and RTSP streams
- Interactive RGB, depth-map, and 3D point-cloud visualization
- Segmentation-assisted measurement and inspection
- Stable Video and Snapshot workflows
- Live and offline reconstruction
- Hardware-aware GPU configuration and runtime profiling
- Isolated model, benchmark, and reconstruction runtimes
MoGe-2 · Depth Pro · UniDepthV2 · Metric3D v2 · Depth Anything V2 · Video Depth Anything · Marigold
Live 3D · Measure & Inspect · Stable Video · Snapshot · Reconstruction · Benchmark
The reconstruction workspace combines classical geometry, learned depth, semantic perception, and neural scene representations while keeping geometry, semantics, measurement, and appearance responsibilities explicit.
| Stage | Method |
|---|---|
| Camera motion & sparse geometry | Visual SLAM and PyCOLMAP Structure-from-Motion |
| Dense geometry | Multi-view depth projection, filtering, and fusion |
| Learned depth support | Monocular depth providers and Depth Anything 3 supporting geometry |
| Semantic geometry | Segmentation-driven object point extraction and 3D geometry |
| Surface representation | Persistent canonical geometry and surface-mesh workflows |
| Neural appearance | Gaussian Splatting and NeRF / Nerfacto |
| Persistence | Reopenable reconstruction sessions with camera, scale, and provenance metadata |
Capture / Images → Camera poses & sparse map → Depth / multi-view geometry → Dense fusion → Canonical scene
Semantic objects and measurements are registered against the scene rather than replacing its canonical geometry. Gaussian Splatting and NeRF/Nerfacto are treated as appearance representations, not authoritative metric geometry.
Measurements remain engineering estimates unless qualified calibration or reference evidence establishes metric scale.
- Ubuntu 24.04 LTS
- Python 3.11 or 3.12
- NVIDIA GPU with a CUDA-compatible PyTorch installation recommended for accelerated inference
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"Optional model runtimes are installed independently through the files under requirements/.
Example:
python -m pip install -r requirements/models-moge2.txt./scripts/run.shor:
python main.pypython scripts/check_hardware.py --assert-ready --require-cudaThe runtime uses CUDA availability, GPU memory, and hardware capability to select an appropriate execution profile.
Python · PyQt6 · PyTorch · CUDA · OpenCV · VisPy · PyCOLMAP · Ultralytics · SAM
Current application version: v0.9.0
The project is under active development.

