ConcertoPretrainingSample<T>
Models & Types · AiDotNet.ComputerVision.Segmentation.PointCloud
One point cloud together with the camera views paired to it, forming a single training example
for Pretrain.
For Beginners
Concerto learns without labels by comparing two things that describe the same scene: the point cloud itself, and photographs taken of it. Each sample here is one scene — the points, where they sit in the world, and the pictures with the camera information needed to work out which pixel each point falls on.
How It Works
PointCoordinates is deliberately separate from Input. The network
consumes whatever tensor layout its encoder expects, while the cross-modal objective needs true
world coordinates to project points into each camera — the two are not interchangeable, and
conflating them silently produces projections that land nowhere.
Constructors
| Constructor | Summary |
|---|---|
ConcertoPretrainingSample() |
Properties
| Property | Summary |
|---|---|
IReadOnlyList<ConcertoPairedView<T>> Views { get; set; } |
Camera views paired with this point cloud. |
Tensor<T> Input { get; set; } |
The tensor fed to the network, in the layout its encoder expects. |
Tensor<T> PointCoordinates { get; set; } |
World-space coordinates of each point, shape [points, 3]. |