Point clouds
Filtering, spatial search, features, segmentation, registration, voxelization, and transforms.
Rust-native spatial intelligence
Capture. Understand. Reconstruct. Act.
A practical guide to SpatialRust's point-cloud, vision, camera, mapping, scene, GPU, and AI capabilities—backed by the complete generated Rust API.
Choose your starting point
Vision and RGB-D
Image processing, local features, multiview geometry, stereo, odometry, and depth-to-world bridges.
Scenes and mapping
Trajectories, pose graphs, TSDF fusion, mesh extraction, surfels, and Gaussian rendering.
GPU and AI
Explicit wgpu image/point kernels, tensor bridges, DLPack, ONNX Runtime, and bounded execution.
Explicit data movement
typed host data→ named upload →device-resident work→ requested readback →owned result
CPU APIs never hide a GPU transfer. Heavy runtimes such as ONNX, ROS 2, CUDA, and codecs stay behind dedicated features and crates.
Architecture at a glance
One data model from capture to action. Every arrow below is a real crate boundary, not an aspirational layer.
flowchart LR
classDef stage fill:#172842,stroke:#28405f,color:#e8eef8;
subgraph Capture["Capture"]
IO["image-io
codecs, metadata"]
CAM["camera
calibration, RGB-D"]
end
subgraph Understand["Understand"]
VIS["vision
imgproc, feature2d, geometry"]
GPU["gpu
explicit wgpu kernels"]
AI["tensor + ai
DLPack, ONNX Runtime"]
end
subgraph Reconstruct["Reconstruct"]
MAP["mapping
trajectories, pose graph"]
SCENE["scene
TSDF, mesh, Gaussian"]
SEM["semantic
embeddings, search"]
end
subgraph Act["Act"]
RUN["runtime
bounded execution graph"]
DIST["distribute
partitions, named transfers"]
end
IO --> VIS
CAM --> VIS
VIS --> GPU
VIS --> AI
VIS --> MAP
AI --> MAP
GPU --> SCENE
MAP --> SCENE
SCENE --> SEM
SEM --> RUN
MAP --> RUN
RUN --> DIST
class IO,CAM,VIS,GPU,AI,MAP,SCENE,SEM,RUN,DIST stage;
Point-cloud filtering, search, segmentation, and registration sit alongside vision and feed the same mapping/scene stages; they are omitted above only for readability. See the algorithm catalog for the full crate list.
Measured, not implied
VGA → 4K
Canonical vision profiles carry accuracy, allocation, and robust timing evidence.
CPU + GPU
Round-trip and device-resident results are reported separately.
OpenCV oracle
Comparison tooling is optional and is not a production runtime dependency.