Multi-View Reconstruction (MapAnything)
========================================

Reconstruct metric-scale depth maps, camera poses, and 3D point clouds from one or more RGB images using the MapAnything deep learning model. Note that quality of metric-scale is dependent upon closeness of provided data with the training distribution thus the quality can vary for out-of-domain cases.

Input
-----

One image set containing N >= 1 same-resolution RGB images with an attached :ref:`camera calibration data component<cameracalibrationdatacomponent>` providing the 3x3 pinhole intrinsic matrix K.

Output
------

- **Depth maps**: Per-view Z-depth maps (float) at the original image resolution. Values are uniformly scaled by 100 (relative to the model's metric output) so the reconstruction fits the display coordinate system.
- **Point clouds** (optional): Per-view coloured 3D point clouds in world space, unprojected using the model's recovered intrinsics and predicted camera poses, in the same scaled coordinate system as the depth maps.

Description
-----------

The algorithm runs the MapAnything reconstruction model, which jointly predicts per-pixel metric depth, camera-to-world poses, recovered intrinsics, per-pixel confidence, and a binary validity mask from N >= 1 unposed RGB images.

The model internally resizes images to 518x518 for processing and maps outputs back to the original input resolution. Any input resolution is accepted.

Depth maps are unprojected using the model's recovered intrinsics and transformed to world space using the predicted camera poses to produce coloured 3D point clouds. The camera poses follow the OpenCV convention (Z forward, X right, Y down).

Parameters
----------

- **Chunk Size**: Maximum number of views processed in a single inference pass (default 0 = process all views at once). If the input has more views than this value, inference is split into chunks.
- **Chunk Overlap**: Number of overlapping images between consecutive chunks when chunking is active (default 1). Must be smaller than **Chunk Size**.
- **Export Point Clouds**: When enabled (default), produces per-view coloured point clouds in world space.
- **Max Depth Threshold**: If > 0, discards points beyond this depth (in metric model units, before the display scaling is applied) when generating point clouds (default 0, meaning no limit).