Source code for miop.default_pipelines

# Copyright (c) 2025, Maxime Paschoud.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
#
# (http://opensource.org/licenses/BSD-3-Clause)
#
# __author__ = "Maxime Paschoud, ETHZ: CMBM"
#
from pipeline import *
from miop import *



def pipeline_fei(pre_processing: bool = True, tiny: bool = False, depth: int = 7):
    
    """
    Create a pipeline for processing FEI TIFF images with metadata.

    This pipeline is designed to process `.tif` images that contain FEI metadata. It includes dense feature
    matching (RoMa), camera movement estimation from metadata, 3D reconstruction, and surface reconstruction.

    Parameters
    ----------
    pre_processing : bool, optional
        If True, the pipeline starts with the `ImageCollection` module to handle image loading and preprocessing.
        Set to False if images are preprocessed manually. Default is True.
    tiny : bool, optional
        If True, use the tiny variant of the RoMa model (faster, less precise). Default is False.
    depth : int, optional
        Depth parameter for Poisson Surface Reconstruction. Higher values yield more detail but consume more memory
        and may produce holes. Default is 7.

    Returns
    -------
    DAG
        A pipeline (DAG object) that can be run with the `execute(inputs)` method.
    """

    img_collection = ImageCollection(metadata_type='fei')
    mat_from_metadata = CameraMvmtFromMetadata()
    dense = DenseRoma(tiny=tiny, num_matches=5000)
    reg = RegistrationByICP()
    rec = Reconstruction(by_pair=False)
    surface_rec = PoissonSurfaceReconstruction(depth=depth)

    img_collection >> mat_from_metadata >> rec >> reg >> surface_rec
    img_collection >> dense >> rec

    list_of_modules = [mat_from_metadata, dense, rec, reg, surface_rec]
    
    if pre_processing:
        list_of_modules.insert(0, img_collection)
        start = img_collection
    else:
        start = [mat_from_metadata, dense]

    stop = surface_rec
    return DAG(start, stop) #, list_of_modules




def pipeline_no_metadata(pre_processing: bool = True, tiny: bool = False, depth: int = 7, threshold: float = 1., batch_size: int = 4):

    """
    Create a pipeline for images without metadata.

    This pipeline is designed for `.tif` images that lack metadata. It includes dense feature matching (RoMa),
    disparity-based point propagation, reprojection error filtering, camera pose estimation, and reconstruction.

    Parameters
    ----------
    pre_processing : bool, optional
        If True, the pipeline starts with the `ImageCollection` module for image loading. Default is True.
    tiny : bool, optional
        If True, use the tiny variant of the RoMa model (faster, less precise). Default is False.
    depth : int, optional
        Depth parameter for Poisson Surface Reconstruction. Higher values yield more detail but consume more memory
        and may produce holes. Default is 7.
    threshold : float, optional
        Threshold (in pixels) for reprojection error. Matches exceeding this error are discarded. Default is 1.0.
    batch_size : int, optional
        Number of images to consider at once for camera motion estimation. Default is 4.

    Returns
    -------
    DAG
        A pipeline (DAG object) that can be run with the `execute(inputs)` method.
    """

    img_collection = ImageCollection(metadata_type='none')
    dense = DenseRoma(tiny=tiny, num_matches=5000)
    inter = Interpolate()
    reg = RegistrationByICP()
    surface_rec = PoissonSurfaceReconstruction(depth=depth)
    corresp = CorrespondingPointsFromDisparity()
    fac = Factorization(threshold, 0, max_matches=10000, batch_size=batch_size)
    rec = Reconstruction(by_pair=False)

    img_collection >> dense >> inter >> corresp >> fac >> rec >> reg >> surface_rec
    dense >> rec

    if pre_processing:
        start = img_collection
    else:
        start = dense

    stop = surface_rec
    return DAG(start, stop)




def pipeline_raft(pre_processing: bool = True, depth: int = 7):
    
    """
    Create a pipeline for FEI TIFF images using RAFT flow.

    This pipeline processes `.tif` images with FEI metadata using RAFT optical flow instead of RoMa
    for dense correspondence estimation. Includes reconstruction and surface generation.

    Parameters
    ----------
    pre_processing : bool, optional
        If True, the pipeline starts with the `ImageCollection` module to load and preprocess images.
        Set to False if preprocessing is already handled. Default is True.
    depth : int, optional
        Depth parameter for Poisson Surface Reconstruction. Higher values yield more detail but consume more memory
        and may produce holes. Default is 7.

    Returns
    -------
    DAG
        A pipeline (DAG object) that can be run with the `execute(inputs)` method.
    """

    img_collection = ImageCollection(metadata_type='fei')
    mat_from_metadata = CameraMvmtFromMetadata()
    raft = RaftFlow()
    corresp = CorrespondingPointsFromDisparity(n_points=10000)
    reg = RegistrationByICP()
    rec = Reconstruction(by_pair=True)
    surface_rec = PoissonSurfaceReconstruction(depth=depth)

    img_collection >> mat_from_metadata >> rec >> reg >> surface_rec
    img_collection >> raft >> corresp >> rec

    if pre_processing:
        start = img_collection
    else:
        start = [mat_from_metadata, raft]

    stop = surface_rec
    return DAG(start, stop)