ConcertoOptions
Options & Configuration · AiDotNet.ComputerVision.Segmentation.PointCloud
Configuration options for Concerto, defaulting to the values published in “Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations” (arXiv:2510.23607).
Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SegmentationModelOptions → ConcertoOptions
For Beginners
The model learns about 3D scenes without any human labels, by (a) checking that two different views of the same point cloud produce consistent features, and (b) checking that the 3D features agree with what a strong 2D image model sees at the same physical locations. Only afterwards is a small labelled head trained on top.
How It Works
Concerto is a SELF-SUPERVISED PRETRAINING method, not a supervised segmentation model. It combines 3D intra-modal self-distillation — a Point Transformer V3 student matched against a momentum-updated teacher — with 2D-3D cross-modal joint embedding against a frozen DINOv2 image encoder. Segmentation is a downstream probe on the learned representation, not the training objective.
Constructors
| Constructor | Summary |
|---|---|
ConcertoOptions() |
Initializes a new instance with the paper’s published defaults. |
ConcertoOptions(ConcertoOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
ConcertoModelSize ModelSize { get; set; } |
Gets or sets which published size variant of the model to build. |
double CrossModalLossWeight { get; set; } |
Weight on the 2D-3D cross-modal joint-embedding term (cosine similarity). |
double IntraModalLossWeight { get; set; } |
Weight on the 3D intra-modal self-distillation term (online clustering cross-entropy). |
double LearningRate { get; set; } |
Peak learning rate for the AdamW optimizer with cosine annealing. |
double TeacherMomentum { get; set; } |
Momentum for the exponential-moving-average update of the teacher encoder. |
double VisibilityDepthToleranceMeters { get; set; } |
Depth agreement, in metres, required for a projected point to count as visible in an image. |
int CrossModalUpcastLevel { get; set; } |
Decoder upcast level at which the cross-modal cosine loss is applied. |
int ImagesPerPointCloud { get; set; } |
Number of images paired with each point cloud during pretraining. |
int IntraModalUpcastLevel { get; set; } |
Decoder upcast level at which the intra-modal clustering loss is applied. |
int PretrainingEpochs { get; set; } |
Number of self-supervised pretraining epochs. |
Methods
| Method | Summary |
|---|---|
void Validate() |
Throws when a value on this instance cannot produce a working model. |