Computer Vision

337 public types in this namespace, organized by kind.

Models & Types (194)

Type Summary
ABCNet<T> ABCNet: single-shot scene text spotting that regresses a cubic Bezier boundary per text instance and rectifies each instance with BezierAlign before recognizing it.
AConvBlock<T> YOLOv9’s AConv downsample: a 2x2 stride-1 average pool, then a 3x3 stride-2 Conv.
ADownBlock<T> YOLOv9’s ADown downsample: a 2x2 stride-1 average pool, then the channels split in two: one half through a 3x3 stride-2 Conv, the other through a 3x3 stride-2 max pool and a 1x1 Conv.
AnatomicalStructure Metadata for a single segmented anatomical structure.
AnchorGenerator<T> Generates anchor boxes for object detection models.
AnchorMatchResult<T> Result of anchor-to-ground-truth matching.
AnchorMatcher<T> Matches anchor boxes to ground truth boxes for training object detectors.
BatchDetectionResult<T> Represents a batch of detection results for multiple images.
BiFPN<T> Bidirectional Feature Pyramid Network (BiFPN) with weighted feature fusion.
BiomedParse<T> BiomedParse: Biomedical image parsing with text prompts.
BitmapFont<T> A simple 5x7 bitmap font for rendering text on images.
ByteTrack<T> ByteTrack: Multi-Object Tracking by Associating Every Detection Box.
C2PSABlock<T> YOLO11’s C2PSA: split in two, run depth PSABlocks on one half, and fuse both halves with a 1x1 Conv.
C2fBlock<T> YOLOv8’s C2f (Jocher et al.
C3k2Block<T> YOLO11’s C3k2: the C2f layout (split, every inner block’s output concatenated), with C3k inner blocks when c3k is set and plain bottlenecks otherwise.
C3kBlock<T> YOLO11’s C3k: a CSP block (C3) of depth 3x3 bottlenecks over half the width, concatenated with a 1x1 bypass.
CATSeg<T> CAT-Seg: Cost Aggregation for open-vocabulary semantic segmentation.
CIoULoss<T> Complete Intersection over Union (CIoU) loss for bounding box regression.
CRAFT<T> CRAFT (Character Region Awareness for Text) detector.
CRNN<T> CRNN (Convolutional Recurrent Neural Network) for text recognition.
CSPBlock<T> Cross-Stage Partial block used in CSP-Darknet.
CSPBottleneckBlock<T> Bottleneck block with residual connection used inside CSP blocks.
CSPDarknet<T> CSP-Darknet backbone network used in YOLO family models (v5, v7, v8).
CUPS<T> CUPS: Scene-Centric Unsupervised Panoptic Segmentation.
CascadeRCNN<T> Cascade R-CNN - Multi-stage object detection with progressive refinement.
CombinedMaskLoss<T> Combined mask loss using BCE + Dice.
ConcertoPairedView<T> One calibrated image paired with a point cloud, supplying the inputs Concerto’s cross-modal joint-embedding objective needs (arXiv:2510.23607).
ConcertoPretrainingSample<T> One point cloud together with the camera views paired to it, forming a single training example for Pretrain.
Concerto<T> Concerto: Hybrid Mamba-Transformer backbone for 3D point clouds.
DBNet<T> DBNet (Differentiable Binarization Network) text detector.
DETRSetLoss<T> DETR-family set prediction loss with exact Hungarian assignment.
DETR<T> DETR (DEtection TRansformer) - End-to-end object detection with transformers.
DEVA<T> DEVA: Tracking Anything with Decoupled Video Segmentation.
DINO<T> DINO (DETR with Improved deNoising anchOr boxes).
DIoULoss<T> Distance Intersection over Union (DIoU) loss for bounding box regression.
DIoUNMS<T> Implements Distance-IoU based Non-Maximum Suppression for improved localization.
DeepSORT<T> DeepSORT (Deep SORT) tracking with appearance features.
DetectionResult<T> Contains the results of object detection on an image.
DetectionStatistics<T> Statistics about detection results.
DetectionTrainingBatch<T> Owns an immutable, unpadded target list for every image in a training batch.
DetectionTrainingTarget<T> A foreground label and a normalized center-format box used to train a detector.
DetectionVisualizer<T> Visualizes object detection results on images.
Detection<T> Represents a single detected object.
DiffCutSegmentation<T> DiffCut: Diffusion-based zero-shot segmentation via graph cuts.
DiffCut<T> DiffCut: Catalyzing Zero-Shot Semantic Segmentation with Diffusion Features and Recursive Normalized Cut.
DiffSeg<T> DiffSeg: Unsupervised Semantic Segmentation from Diffusion Model Self-Attention Maps.
DocumentBlock<T> A block of text in a document (paragraph, header, etc.).
DocumentBlock<T> A block in a document (paragraph, table, figure, etc.).
DocumentLayoutResult<T> Result of document layout analysis.
DocumentLine<T> A line of text in a document.
DocumentOCRResult<T> Result of document OCR.
DocumentReader<T> Document reader for OCR with layout analysis.
EAST<T> EAST (Efficient and Accurate Scene Text) detector.
ELAN1Block<T> YOLOv9’s ELAN1 (GELAN-t/s stage 2): RepNCSPELAN4’s layout with plain 3x3 Convs in place of the RepNCSP stages.
EdgeSAM<T> EdgeSAM: Prompt-in-the-Loop Distillation for SAM on edge devices.
EfficientNet<T> EfficientNet backbone for efficient feature extraction.
EfficientSAM<T> EfficientSAM: Leveraged Masked Image Pretraining for efficient SAM.
EfficientTAM<T> EfficientTAM (Efficient Track Anything): SAM 2’s promptable video tracking on a lightweight plain ViT, for devices.
EoMT<T> EoMT: Encoder-only Mask Transformer for universal image segmentation.
FPN<T> Feature Pyramid Network (FPN) for multi-scale feature fusion.
FastSAM<T> FastSAM: Fast Segment Anything Model based on YOLOv8-Seg.
FasterRCNN<T> Faster R-CNN - Two-stage object detection with region proposal network.
GIoULoss<T> Generalized Intersection over Union (GIoU) loss for bounding box regression.
GroundingToken A grounding token linking a text span to an image region.
InstanceMask<T> Represents a single instance with bounding box and segmentation mask.
InstanceSegmentationResult<T> Result of instance segmentation.
InstanceSegmentationTrainingTarget<T> One object to learn: its class and box, and its binary mask over the whole input image.
InternImage<T> InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions.
KMaXDeepLab<T> kMaX-DeepLab: k-means Mask Transformer for panoptic segmentation.
KoLeoRegularizer<T> The Kozachenko-Leonenko entropy regulariser of DINOv2 (Sablayrolles et al.
LISA<T> LISA: Reasoning Segmentation via Large Language Model.
Mask2Former<T> Mask2Former: Masked-attention Mask Transformer for Universal Image Segmentation.
MaskAdapter<T> Mask-Adapter: Adding SAM to open-vocabulary segmentation via mask prediction.
MaskBCELoss<T> Binary Cross-Entropy loss for mask prediction.
MaskDINO<T> Mask DINO: Towards A Unified Transformer-based Framework for Object Detection and Segmentation.
MaskDiceLoss<T> Dice loss for mask prediction.
MaskFocalLoss<T> Focal loss for mask prediction (addresses class imbalance).
MaskHead<T> Mask prediction head for instance segmentation.
MaskRCNN<T> Mask R-CNN for instance segmentation.
MaskVisualizer<T> Visualizes instance segmentation results on images.
MedNeXt<T> MedNeXt: Transformer-driven scaling of ConvNets for medical segmentation.
MedSAM2<T> MedSAM2: SAM 2.1 fine-tuned to segment anything in 3D medical images and videos.
MedSAM<T> MedSAM: Segment Anything in Medical Images.
MedSegDiffV2Segmentation<T> MedSegDiff-V2 Segmentation: Diffusion-based medical image segmentation pipeline.
MedSegDiffV2<T> MedSegDiff-V2: Spectrum-space diffusion for medical segmentation.
MedicalSAM2<T> Medical SAM 2: SAM 2 for medical images and volumes, segmenting sets of 2D images as pseudo-videos through a self-sorting memory bank.
MedicalSegmentationOutput<T> Output for medical image segmentation with volumetric support and clinical metadata.
MixedQueryTransformer<T> MixedQueryTransformer (MQ-Former): Dynamic Query Melding for Multi-Dataset Segmentation.
MobileSAM<T> MobileSAM: Faster Segment Anything with TinyViT encoder.
NMS<T> Implements Non-Maximum Suppression (NMS) algorithms for removing duplicate detections.
NnUNet<T> nnU-Net: Self-configuring framework for medical image segmentation.
OCRResult<T> Result of OCR processing.
OCRVisualizer<T> Visualizes OCR (Optical Character Recognition) results on images.
ODISESegmentation<T> ODISE Segmentation: Panoptic segmentation via diffusion model features.
ODISE<T> ODISE: Open-vocabulary DIffusion-based panoptic SEgmentation.
OMGLLaVA<T> OMG-LLaVA: Bridging Image-Level, Object-Level, and Pixel-Level understanding.
OMGSeg<T> OMG-Seg: Is One Model Good Enough For All Segmentation?
ObjectTrackSummary Summary of an object’s track across a video sequence.
OneFormer<T> OneFormer: One Transformer to Rule Universal Image Segmentation.
OpenVocabSAM<T> Open-Vocabulary SAM: SAM with text-based open-vocabulary recognition.
OpenVocabSegmentationOutput<T> Output for open-vocabulary and referring segmentation models.
PANet<T> Path Aggregation Network (PANet) for enhanced multi-scale feature fusion.
PIDNet<T> PIDNet: A Real-time Semantic Segmentation Network Inspired by PID Controllers.
PSABlock<T> YOLO11’s PSABlock: multi-head self-attention with a depthwise positional encoding on the values, then a 2x FFN, each added back to its input.
PixelLM<T> PixelLM: Pixel Reasoning with Large Multimodal Model.
PointCloudSegmentationOutput<T> Output for 3D point cloud segmentation.
PointTransformerV3Encoder<T> The Point Transformer V3 backbone (Wu et al., CVPR 2024), shared by PointTransformerV3 and Sonata: serialization, a sparse-convolution or linear stem, staged blocks of serialized patch attention with xCPE, grid pooling between stages an…
PointTransformerV3<T> Point Transformer V3: Simpler, Faster, Stronger – per-point semantic segmentation of 3D point clouds.
PrototypeMaskHead<T> Prototype-based mask head for YOLO and SOLOv2.
RPN<T> Region Proposal Network (RPN) - Generates object proposals for two-stage detectors.
RTDETR<T> RT-DETR (Real-Time DEtection TRansformer) - the first real-time end-to-end object detector.
RecognizedText<T> Represents recognized text in an image region.
RepBottleneckBlock<T> YOLOv9’s RepNBottleneck: RepConvN then a 3x3 Conv, with the input added back when shortcut is set.
RepCSPBlock<T> YOLOv9’s RepNCSP: a CSP (C3) block of depth RepN bottlenecks over half the width, concatenated with a 1x1 bypass.
RepConvNBlock<T> YOLOv9’s RepConvN at training time: a 3x3 Conv and a 1x1 Conv (each with batch norm, no activation) summed, then SiLU.
RepNCSPELAN4Block<T> YOLOv9’s RepNCSPELAN4 (GELAN): a 1x1 Conv split in two, two RepNCSP + 3x3 Conv stages chained on the second half, and a 1x1 Conv over all four maps.
RepViTSAM<T> RepViT-SAM: Real-time SAM with RepViT backbone.
ResNet<T> ResNet backbone network for feature extraction.
SAM21<T> SAM 2.1: Segment Anything Model 2.1, Meta’s refreshed SAM 2 for images and videos.
SAMHQ<T> SAM-HQ: Segment Anything in High Quality.
SAM<T> Segment Anything Model (SAM): the first promptable foundation model for image segmentation.
SAN<T> SAN: Side Adapter Network for open-vocabulary semantic segmentation.
SED<T> SED: A Simple Encoder-Decoder for open-vocabulary semantic segmentation.
SEEM<T> SEEM: Segment Everything Everywhere All at Once.
SOLOv2<T> SOLOv2 (Segmenting Objects by Locations v2) for instance segmentation.
SORT<T> SORT (Simple Online and Realtime Tracking) implementation.
SPPELANBlock<T> YOLOv9’s SPPELAN: a 1x1 Conv, three chained 5x5 stride-1 max pools, and a 1x1 Conv over the four concatenated maps.
SceneTextReader<T> End-to-end scene text reader that combines detection and recognition.
SegFormer<T> SegFormer: Simple and Efficient Semantic Segmentation with Transformers.
SegGPT<T> SegGPT: Segmenting Everything In Context via in-context learning.
SegMamba<T> SegMamba: long-range sequential modeling Mamba for 3D medical image segmentation (Xing et al., 2024, arXiv:2401.13560).
SegNeXt<T> SegNeXt: Rethinking Convolutional Attention Design for Semantic Segmentation.
SegmentInfo<T> Metadata for a single segment in panoptic/instance output.
SegmentationEvaluation Comprehensive evaluation metrics for segmentation models covering all standard benchmarks.
SegmentationMask<T> Represents a single segmentation mask with associated metadata.
SegmentationOutput<T> Unified output type for all segmentation tasks, combining semantic, instance, and panoptic results.
SegmentationPrompt<T> Represents a user prompt for interactive/promptable segmentation models.
SegmentationSample<T> Represents a segmentation dataset sample with image, masks, and metadata.
SinkhornKnoppCentering<T> Sinkhorn-Knopp centering (Caron et al.
SlimSAM<T> SlimSAM: Pruned and distilled SAM for efficient segmentation.
SoftNMS<T> Implements Soft-NMS algorithm which reduces confidence of overlapping boxes instead of removing them.
SonataClusterCrossEntropyLoss<T> Sonata’s cluster cross-entropy: -sum(target * log_softmax(student / tau)) per match, averaged within each scene and then across scenes (segment_coo(..., reduce="mean").mean()).
Sonata<T> Sonata: self-supervised learning of reliable point representations (Wu et al., CVPR 2025) – a Point Transformer V3 encoder pretrained by point self-distillation, read out by a linear probe.
SpaceFillingCurveSet The ordered serialization curves of a Point Transformer V3 encoder, saved by registered name.
SwinTransformer<T> Swin Transformer backbone for hierarchical vision transformer feature extraction.
SwinUNETR<T> Swin UNETR: Swin Transformer encoder for 3D medical segmentation.
TaskAlignedDetectionLoss<T> Task-aligned assignment with BCE, CIoU and distribution focal losses for anchor-free YOLO heads.
TensorData Represents tensor data extracted from a weight file.
TextDetectionResult<T> Result of text detection containing detected text regions.
TextDetectionTrainingBatch The text polygons of every image in a training batch.
TextPolygonTarget One annotated text region: a polygon in source-image pixels and, optionally, its transcription.
TextRegion<T> Represents a detected text region in an image.
TrOCR<T> TrOCR (Transformer-based OCR) for text recognition.
Track<T> Represents a tracked object across frames.
TrackingResult<T> Result of tracking on a frame.
TransUNet<T> TransUNet: Transformers make strong encoders for medical segmentation.
TwoStageDetectionLoss<T> Region proposal and region-of-interest losses for Faster R-CNN and Cascade R-CNN.
U2Seg<T> U2Seg: Unified Unsupervised Segmentation framework.
UMamba<T> U-Mamba: Hybrid CNN-Mamba architecture for medical segmentation.
UNINEXT<T> UNINEXT: Universal Instance Perception as Object Discovery and Retrieval.
UniVS<T> UniVS: Unified and Universal Video Segmentation.
UniverSeg<T> UniverSeg: Universal Medical Image Segmentation via cross-attention.
VGG16BNBackbone<T> VGG-16 with batch normalization as a feature-pyramid backbone, cut the way CRAFT cuts it.
VGG16Backbone<T> VGG-16 (configuration D, no batch normalization) as a feature-pyramid backbone, cut the way EAST cuts it.
VMamba<T> VMamba: Visual State Space Model with Cross-Scan for 2D understanding.
ViMUNet<T> ViM-UNet: Vision Mamba for medical image segmentation in U-Net.
ViTAdapter<T> ViT-Adapter: Vision Transformer Adapter for Dense Predictions.
ViTCoMer<T> ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction.
VideoFrameSegmentation<T> Segmentation result for a single video frame.
VideoLISA<T> Video-LISA: Language-instructed video segmentation with reasoning.
VideoSegmentationOutput<T> Output for video segmentation containing per-frame masks with temporal tracking.
VisionMamba<T> Vision Mamba (Vim): Bidirectional State Space Model for vision.
WeightDownloader Downloads and caches pre-trained model weights from remote URLs.
WeightLoader Loads pre-trained model weights from various file formats.
XDecoder<T> X-Decoder: Generalized Decoding for Pixel, Image, and Language.
YOLO11Seg<T> YOLO11-Seg: Ultralytics next-generation real-time instance segmentation.
YOLO26Seg<T> YOLO26-Seg: Latest YOLO family instance segmentation with advanced CSP design.
YOLOSeg<T> YOLOv8-Seg for instance segmentation.
YOLOv10<T> YOLOv10 object detector with NMS-free training and consistent dual assignments.
YOLOv11Backbone<T> YOLO11’s backbone (ultralytics yolo11.yaml): Conv/C3k2 stages, SPPF, then C2PSA attention on P5.
YOLOv11Neck<T> YOLO11’s PAN-FPN neck: YOLOv8’s topology with C3k2 in place of C2f.
YOLOv11<T> YOLOv11 object detector with enhanced feature extraction and attention mechanisms.
YOLOv12Seg<T> YOLOv12-Seg: Attention-centric real-time instance segmentation.
YOLOv8Backbone<T> YOLOv8’s backbone: Conv stem, four stride-2 Conv + C2f stages, then SPPF (Jocher et al.
YOLOv8Neck<T> YOLOv8’s PAN-FPN neck: top-down nearest upsampling with concatenation and C2f (no shortcut), then bottom-up stride-2 Conv with concatenation and C2f.
YOLOv8Seg<T> YOLOv8-Seg: Ultralytics real-time instance segmentation.
YOLOv8<T> YOLOv8 object detector - anchor-free, decoupled head architecture.
YOLOv9Backbone<T> YOLOv9’s backbone (Wang et al.
YOLOv9Neck<T> YOLOv9’s head (PAN-FPN): SPPELAN on P5, top-down upsample + concat + RepNCSPELAN4, then bottom-up AConv/ADown + concat + RepNCSPELAN4, with every size’s own widths.
YOLOv9Seg<T> YOLOv9-Seg: Instance segmentation with Programmable Gradient Information.
YOLOv9<T> YOLOv9 object detector: the GELAN architecture with Programmable Gradient Information (PGI).
YoloBottleneck<T> YOLO’s Bottleneck: two 3x3 Convs, with the input added back when shortcut and widths match.
YoloConv<T> YOLO’s Conv: convolution, batch normalization, then SiLU (or no activation when act is false).
YoloPanBlocks<T> The six blocks of a YOLO PAN-FPN neck in execution order, and the width of each output level.

Layers (1)

Type Summary
SPPFLayer<T> Spatial Pyramid Pooling - Fast: a 1x1 Conv to half width, three chained 5x5 stride-1 max pools, and a 1x1 Conv over the four concatenated maps (equivalent to SPP’s 5/9/13 pools, at a third of the cost).

Base Classes (18)

Type Summary
InstanceSegmentationBase<T> Abstract base class for instance segmentation models that detect and mask individual object instances.
InstanceSegmenterBase<T> Base class for instance segmentation models.
MedicalSegmentationBase<T> Abstract base class for medical image segmentation models handling 2D slices and 3D volumes.
NeckBase<T> Base class for neck modules that perform multi-scale feature fusion.
OCRBase<T> Base class for OCR models.
ObjectDetectorBase<T> Base class for all object detection models.
ObjectTrackerBase<T> Base class for object trackers.
OpenVocabSegmentationBase<T> Abstract base class for open-vocabulary segmentation models that segment objects described by arbitrary text without being limited to a fixed class set.
PanopticSegmentationBase<T> Abstract base class for panoptic segmentation models that unify semantic and instance segmentation.
PromptableSegmentationBase<T> Abstract base class for promptable segmentation models like SAM that accept user prompts (points, boxes, masks) to segment specific objects.
ReferringSegmentationBase<T> Abstract base class for referring segmentation models that segment objects from natural language descriptions with complex reasoning about spatial relationships and attributes.
SegmentationModelBase<T> Abstract base class for all segmentation models, providing common dual-mode (native + ONNX) infrastructure, batch handling, forward/backward passes, and serialization.
SemanticSegmentationBase<T> Abstract base class for semantic segmentation models that assign a class label to every pixel.
TextDetectorBase<T> Base class for text detection models.
VideoSegmentationBase<T> Abstract base class for video segmentation models that track and segment objects across frames.
VisionTaskModelBase<T> The training and state plumbing the vision task families share: object detectors, text detectors and text recognisers (ObjectDetectorBase, TextDetectorBase, OCRBase).
YoloPanNeckBase<T> The PAN-FPN neck shared by YOLOv8 and YOLO11: top-down nearest upsampling with concatenation and a fusion block, then bottom-up stride-2 Conv with concatenation and a fusion block.
YoloStagedBackboneBase<T> The staged YOLO backbone shared by YOLOv8 and YOLO11: a chain of stages with P3, P4 and P5 tapped at fixed stage indices.

Interfaces (4)

Type Summary
ISonataDistillationLoss<T> The per-match distillation loss between student scores and teacher targets (Sonata uses the cluster cross-entropy, SonataClusterCrossEntropyLoss).
ISonataFeatureRegularizer<T> An extra regulariser on the student’s point embeddings, added to the distillation loss (for example the KoLeo spreading term, KoLeoRegularizer).
ISonataTargetCentering<T> Turns the teacher’s prototype scores into the soft assignment targets the student is trained to predict (Sonata uses Sinkhorn-Knopp centering, SinkhornKnoppCentering).
ISpaceFillingCurve A space-filling curve that maps a 3D integer grid coordinate to a position on a 1D curve.

Enums (18)

Type Summary
DecayMethod<T> Soft-NMS decay method.
DocumentBlockType Type of document block.
DocumentBlockType Types of document blocks.
InstanceSegmentationArchitecture Instance segmentation architectures.
MatchLabel Label indicating how an anchor should be treated during training.
MedicalSAM2ModelSize The SAM 2 sizes Medical SAM 2 releases configurations for.
MemoryBankResampling How a self-sorting memory bank draws the memories a frame attends to.
MemoryBankUpdateRule How a self-sorting memory bank decides what to keep.
MemoryCrossAttention How the memory attention’s cross-attention reads the spatial memories.
OCRMode OCR processing modes.
PromptType Type of segmentation prompt.
SegmentationTaskType Type of segmentation task.
SwinVariant Swin Transformer variant enumeration.
TextDetectionArchitecture Text detection architecture types.
TextDetectionModel Text detection model types.
TextRecognitionModel Text recognition model types.
TextRegionType Type of text region.
TrackState Track lifecycle states.

Structs (2)

Type Summary
CubicBezierCurve<T> A cubic Bezier curve over four control points, and the least-squares fit that recovers those control points from annotated polygon vertices.
TextInstance<T> A spotted text instance: where it is, how confident, and what it says.

Options & Configuration (90)

Type Summary
BiomedParseOptions Configuration options for BiomedParse biomedical foundation segmentation.
ByteTrackOptions<T> Association thresholds of ByteTrack; the defaults are those of the official BYTETracker (Zhang et al.
CATSegOptions Configuration options for CAT-Seg cost aggregation open-vocabulary segmentation.
CSPDarknetOptions Configuration options for the CSPDarknet detection backbone.
CUPSOptions Configuration options for CUPS unsupervised panoptic segmentation.
ConcertoOptions Configuration options for Concerto, defaulting to the values published in “Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations” (arXiv:2510.23607).
DEVAOptions Configuration options for DEVA decoupled video segmentation.
DetrSetLossOptions Weights and classification form for a DETR-family set prediction loss.
DiffCutOptions Configuration options for the DiffCut semantic segmentation model.
DiffCutSegmentationOptions Configuration options for DiffCut diffusion-based zero-shot segmentation.
DiffSegOptions Configuration options for the DiffSeg semantic segmentation model.
EdgeSAMOptions Configuration options for EdgeSAM edge-optimized SAM.
EfficientNetBackboneOptions Configuration options for the EfficientNetBackbone detection backbone.
EfficientSAMOptions Configuration options for EfficientSAM (SAMI-pretrained fast SAM).
EfficientTAMOptions Configuration options for EfficientTAM, Efficient Track Anything (Xiong et al.
EoMTOptions Configuration options for the EoMT (Encoder-only Mask Transformer) model.
FastSAMOptions Configuration options for FastSAM (CNN-based fast Segment Anything).
InstanceSegmentationOptions<T> Options for instance segmentation models.
InternImageOptions Configuration options for the InternImage semantic segmentation model.
KMaXDeepLabOptions Configuration options for kMaX-DeepLab panoptic segmentation.
LISAOptions Configuration options for LISA reasoning segmentation.
Mask2FormerOptions Configuration options for the Mask2Former universal segmentation model.
MaskAdapterOptions Configuration options for Mask-Adapter open-vocabulary segmentation.
MaskDINOOptions Configuration options for the Mask DINO model.
MedNeXtOptions Configuration options for MedNeXt medical segmentation.
MedSAM2Options Configuration options for MedSAM2, MedSAM2 (Ma et al.
MedSAMOptions Configuration options for MedSAM universal medical segmentation.
MedSegDiffV2Options Configuration options for MedSegDiff-V2 transformer diffusion medical segmentation.
MedSegDiffV2SegmentationOptions Configuration options for MedSegDiff-V2 transformer diffusion medical segmentation.
MedicalSAM2Options Configuration options for MedicalSAM2, Medical SAM 2 (Zhu et al.
MixedQueryTransformerOptions Configuration options for the MixedQueryTransformer (MQ-Former) model.
MobileSAMOptions Configuration options for MobileSAM (distilled lightweight SAM).
NeckConfig Configuration for neck modules.
NnUNetOptions Configuration options for nnU-Net v2 medical segmentation.
OCROptions<T> Options for OCR models.
ODISEOptions Configuration options for ODISE open-vocabulary panoptic segmentation.
ODISESegmentationOptions Configuration options for ODISE diffusion-based panoptic segmentation.
OMGLLaVAOptions Configuration options for OMG-LLaVA unified pixel-level reasoning.
OMGSegOptions Configuration options for the OMG-Seg model.
OneFormerOptions Configuration options for the OneFormer universal segmentation model.
OpenVocabSAMOptions Configuration options for Open-Vocabulary SAM interactive recognition.
PIDNetOptions Configuration options for PIDNet real-time segmentation.
PixelLMOptions Configuration options for PixelLM pixel-level reasoning segmentation.
PointTransformerV3EncoderOptions The Point Transformer V3 backbone settings shared by PointTransformerV3Options and SonataOptions.
PointTransformerV3Options Configuration options for Point Transformer V3 semantic segmentation.
RepViTSAMOptions Configuration options for RepViT-SAM real-time mobile SAM.
ResNetBackboneOptions Configuration options for the ResNetBackbone detection backbone.
SAM21Options Configuration options for SAM21, SAM 2.1 (Meta’s refreshed SAM 2 checkpoints, with the temporal encoding of the object pointers and the occlusion embedding of the memories).
SAMHQOptions Configuration options for the SAM-HQ (High-Quality Segment Anything) model.
SAMOptions Configuration options for the Segment Anything Model (SAM).
SANOptions Configuration options for SAN (Side Adapter Network) open-vocabulary segmentation.
SEDOptions Configuration options for SED simple encoder-decoder open-vocabulary segmentation.
SEEMOptions Configuration options for SEEM interactive segmentation.
SegFormerOptions Configuration options for the SegFormer semantic segmentation model.
SegGPTOptions Configuration options for SegGPT in-context segmentation.
SegMambaOptions Configuration options for SegMamba 3D volumetric Mamba segmentation.
SegNeXtOptions Configuration options for SegNeXt (Guo et al.
SegmentationDatasetConfig Configuration for a segmentation dataset.
SegmentationVisualizationConfig Visualization settings and utilities for segmentation outputs.
SlimSAMOptions Configuration options for SlimSAM (pruned SAM with 1.4% params).
SonataOptions Configuration options for Sonata, self-supervised point self-distillation of a Point Transformer V3 encoder (Wu et al., CVPR 2025).
SwinTransformerOptions Configuration options for the SwinTransformer detection backbone.
SwinUNETROptions Configuration options for Swin-UNETR brain tumor segmentation.
TaskAlignedLossOptions Assignment and loss settings for task-aligned anchor-free YOLO training.
TextDetectionOptions<T> Options for text detection models.
TrOCROptions<T> TrOCR’s architecture hyperparameters.
TrackingOptions<T> Options for object tracking.
TransUNetOptions Configuration options for TransUNet medical segmentation.
TwoStageDetectionLossOptions Sampling and weighting for training two-stage (R-CNN family) detectors.
U2SegOptions Configuration options for the U2Seg model.
UMambaOptions Configuration options for U-Mamba CNN+Mamba medical segmentation.
UNINEXTOptions Configuration options for the UNINEXT model.
UniVSOptions Configuration options for UniVS universal video segmentation.
UniverSegOptions Configuration options for UniverSeg few-shot medical segmentation.
VGG16BNBackboneOptions Configuration options for the VGG16-BN detection backbone.
VGG16BackboneOptions Configuration options for the VGG16 detection backbone.
VMambaOptions Configuration options for VMamba visual state space model segmentation.
ViMUNetOptions Configuration options for ViM-UNet Vision Mamba + U-Net biomedical segmentation.
ViTAdapterOptions Configuration options for the ViT-Adapter semantic segmentation model.
ViTCoMerOptions Configuration options for the ViT-CoMer semantic segmentation model.
VideoLISAOptions Configuration options for VideoLISA video reasoning segmentation.
VisionMambaOptions Configuration options for Vision Mamba (Vim) bidirectional SSM segmentation.
VisualizationOptions Options for visualization.
XDecoderOptions Configuration options for the X-Decoder model.
YOLO11SegOptions Configuration options for YOLO11-Seg instance segmentation.
YOLO26SegOptions Configuration options for YOLO26-Seg instance segmentation.
YOLOv12SegOptions Configuration options for YOLOv12-Seg instance segmentation.
YOLOv8SegOptions Configuration options for the YOLOv8-Seg instance segmentation model.
YOLOv9SegOptions Configuration options for YOLOv9-Seg instance segmentation.
YoloBackboneOptions Configuration options for the YOLOv8, YOLOv9 and YOLO11 detection backbones.

Helpers & Utilities (10)

Type Summary
AnchorPresets Pre-defined anchor configurations for common models.
BezierAlign BezierAlign: rectifies an arbitrarily curved text instance into a rectangular feature map by sampling along its two boundary Bezier curves.
ConcertoCrossModalObjective<T> Concerto’s 2D-3D cross-modal joint-embedding objective (arXiv:2510.23607).
ConcertoIntraModalObjective<T> Concerto’s 3D intra-modal self-distillation objective (arXiv:2510.23607), built on the Sonata / DINOv2 online-clustering formulation.
PretrainedRegistry Registry of pre-trained model weights with their download URLs.
ResNetBackboneSelection Picks the ResNet a detector builds from its options’ BackboneType.
SegmentationDatasets Configuration for standard segmentation datasets.
SegmentationRenderer Draws segmentation masks over an image, producing an overlay image tensor.
SegmentationTensorOps Static helper methods for common segmentation tensor operations (argmax, softmax, etc.) used by models implementing segmentation interfaces.
SpaceFillingCurves The built-in serialization orders and a name registry for custom ones.