scanlt3d 0.1.4

Realtime 3D (depth) + detection pipeline for laptop cameras (OpenCV optional)

scanlt3d

Realtime camera pipeline for detection + monocular depth ("3D") with pluggable backends.

• Project/package name (PyPI): scanlt3d
• Import name (Python): scanlt

This library does not require OpenCV.

Install

Minimal (always works)

CPU-only, no special hardware acceleration:

pip install scanlt3d

Optional extras

You can install extra backends depending on your machine.

pip install "scanlt3d[onnx]"       # ONNX Runtime (recommended cross-platform)
pip install "scanlt3d[torch]"      # PyTorch (CUDA/MPS support)
pip install "scanlt3d[mediapipe]"  # (planned) camera source via MediaPipe

Note: Hardware acceleration depends on which runtime is installed and available on your system (CUDA/DirectML/MPS).

Quickstart

1) Demo webcam with segmentation mask (recommended)

This is the easiest way to understand what scanlt3d does.

Install OpenCV for preview:

pip install "scanlt3d[opencv]"

Then run:

import scanlt

# Downloads a default segmentation model on first run (profile="fast") and caches it.
# Press `q` to quit.
scanlt.demo_webcam()

Choose a different model profile (auto-downloads if needed):

import scanlt

scanlt.demo_webcam(profile="balanced")
scanlt.demo_webcam(profile="quality")

Notes:

• The first run may take time to download the model.
• If OpenCV is not installed, demo_webcam() will run headless and you should use on_result to consume masks.

2) Low-level loop (always works)

import scanlt

# Runs a realtime loop with a built-in dummy camera source.
# This ensures `import + run()` never fails even if no webcam/backend is available.
scanlt.run()

Use by hardware (CPU / NVIDIA / Windows iGPU / Mac M)

scanlt can auto-detect the best available backend:

import scanlt   

print(scanlt.choose_backend())

1) CPU (Intel/AMD)

Install:

pip install scanlt3d
# recommended runtime
pip install "scanlt3d[onnx]"

What to expect:

• Best compatibility.
• Realtime depends heavily on your detector/depth model sizes.
• Use lower resolution / run depth less frequently for higher FPS.

2) NVIDIA GPU (CUDA)

Two typical options:

• ONNX Runtime CUDA (best for ONNX models)
• PyTorch CUDA (best for torch models)

Install (choose one):

# Option A: PyTorch CUDA
pip install "scanlt3d[torch]"

# Option B: ONNX Runtime (you must install a CUDA-enabled onnxruntime build)
pip install "scanlt3d[onnx]"

Notes:

• If choose_backend() returns cuda, scanlt detected a CUDA-capable runtime.
• CUDA packaging varies by OS/driver; if CUDA runtime is not available, scanLt falls back to CPU.

3) Windows Intel/AMD iGPU (DirectML)

scanlt can pick dml if your ONNX Runtime installation exposes a DirectML provider.

Install:

pip install "scanlt3d[onnx]"

Notes:

• DirectML is Windows-specific.
• If DirectML is not available, scanLt falls back to CPU.

4) macOS Apple Silicon (M1/M2/M3)

Install:

pip install scanlt3d
pip install "scanlt3d[torch]"

Notes:

• scanLt will try to use mps if PyTorch MPS is available.
• If MPS is not available or unsupported by the model, it falls back to CPU.

Force backend (override auto-detect)

You can override backend selection via environment variable:

• SCANLT_BACKEND=cpu
• SCANLT_BACKEND=cuda
• SCANLT_BACKEND=dml
• SCANLT_BACKEND=mps

Example:

# Windows PowerShell
setx SCANLT_BACKEND cuda

(Then restart your terminal.)

Provide your own detector/depth

scanLt is designed to let you plug in your own models.

Minimal interfaces

• Detector.predict(frame) -> list[Detection]
• DepthEstimator.predict(frame, detections=None) -> depth_map

Example

import scanlt

class MyDetector:
    def predict(self, frame):
        # return list of scanlt.api.Detection
        return []

class MyDepth:
    def predict(self, frame, detections=None):
        import numpy as np
        h, w = frame.shape[:2]
        return np.zeros((h, w), dtype=np.float32)


def on_result(res):
    # res.frame: np.ndarray (H,W,3)
    # res.detections: list
    # res.depth: np.ndarray (H,W) or None
    # res.fps: float
    print(res.fps)


scanlt.run(detector=MyDetector(), depth=MyDepth(), on_result=on_result, max_frames=100)

Troubleshooting

pip install succeeds but choose_backend() is still CPU

This usually means:

• CUDA / DirectML / MPS runtime is not installed or not detected
• your model runtime doesn't support the needed execution provider

scanLt will always fall back to CPU to stay usable.

I want real webcam input

Current default run() uses a dummy source. For real camera sources, you will plug in a FrameSource (or enable the MediaPipe source once implemented).