reconstruction

Definition

Recovering an input after it has been compressed, corrupted, or partially hidden — and the recovered output itself. Reconstruction error is the training signal behind autoencoders and masked modelling, and the same word covers 3D reconstruction, where the target is scene geometry inferred from images. It differs from generation in that a ground-truth target exists, which makes the objective easy to define but says little about whether the learned representation is useful.

Explainers using this term

• Paper Walkthrough: Stop Anchoring to Frame One — Scal3R's Multi-Reference Relative Pose QueryScal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction
• Paper Walkthrough: Eleven Frames Are Enough — Revisiting Local Context for Long-Horizon Streaming 3D ReconstructionRevisiting Local Context for Long-Horizon Streaming 3D Reconstruction
• Paper Explained — GenFirst: Let Generation Shape the Latent Space First, Reconstruction SecondGenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling