Skip to content

Models Overview

SpatialHub provides ONNX Runtime adapters for 7 computer vision model architectures across perception tasks. All adapters return standard Python dataclass contracts defined in spatialhub.structures.

Perception Models Summary Table

Model Architecture Task Primary Adapter Class Available Model Options / Presets Supported Sub-Adapters Returned Result Contract
EfficientLoFTR Semi-dense Feature Matching EfficientLoFTR "full" (Default), "opt" N/A MatchResult
Depth Anything 3 Monocular & Multi-View Depth DepthAnything3 "da3_small", "da3_base" (Default), "da3_large", "da3_giant", "da3mono_large", "da3metric_large", "da3nested_*_large" N/A DepthPredictionResult
DINOv2 Image Feature Extraction DINOv2 "vits14", "vitb14", "vitl14" (Default), "vitg14" N/A FeatureExtractionResult
FastSAM Real-Time Proposal Segmentation FastSAM "x" (Default), "s" N/A SegmentationResult
SAM Automatic Mask Generation (AMG) SAM "vit_h" (Default), "vit_l", "vit_b" N/A SegmentationResult
CNOS CAD Zero-Shot Object Segmentation CNOS 3D CAD Mesh (.ply, .obj, .stl) Segmentors: FastSAMAdapter, SAMAdapter
Descriptor: DINOv2Adapter		 SegmentationResult
FoundationPose Model-based 6D Object Pose Estimation & Tracking FoundationPose 3D CAD Mesh (.ply, .obj, .stl) Refiner: PoseRefinePredictor
Scorer: ScorePredictor		 PoseEstimationResult

Detailed Technical References

  • EfficientLoFTR Documentation : "full" and "opt" precision options, dimension alignment math, coordinate scaling, and matching API.
  • Depth Anything 3 Documentation : Main foundation series (small, base, large, giant), Metric series, Monocular series, and Nested dual-model alignment.
  • DINOv2 Documentation : Vision Transformer backbones (vits14, vitb14, vitl14, vitg14), ImageNet normalization, and unit L2 normalization.
  • FastSAM Documentation : YOLOv8-Seg models (x, s), prototype mask decoding, NMS filtering, and proposal generation.
  • SAM Documentation : SAM variants (vit_h, vit_l, vit_b), decoupled Image Encoder / Mask Decoder ONNX execution, point grid sampling, and AMG.
  • CNOS Documentation : 3D CAD mesh template rendering, pluggable segmentors (FastSAM, SAM), DINOv2 descriptor matching, and zero-shot instance segmentation.
  • FoundationPose Documentation : Iterative neural pose refinement (RefineNet), pairwise comparison tournament scoring (ScoreNet), ModernGL G-buffer atlas rendering, and 6D object pose registration/tracking.