Tasks configuration

Introduction

Each task has some default configuration but to create a working pipeline, you need to set some parameters in a TOML configuration file.

There are two types of parameters:

1. those associated with luigi: upstream tasks, requirements & outputs
2. those tuning the used algorithms: variables & parameters

For examples look at the configurations examples provided in plant3dvision/configs.

I/O tasks

These tasks refer to the inputs and outputs of each algorithmic tasks. They are often related to task defined in the plantdb library.

FilesetExists

Check that the given Fileset id exists (in dataset).

No luigi parameters (no upstream task).

List of task variables:

• fileset_id: the id (str) to check for existence.

ImagesFilesetExists(FilesetExists)

Check that the image file set exists.

No luigi parameters (no upstream task).

List of task variables:

• fileset_id: the id (str) to check for existence 'images' by default.

ModelFileset(FilesetExists)

Check that the trained ML models file set exists.

No luigi parameters (no upstream task).

List of task variables:

• fileset_id: the id (str) to check for existence 'models' by default.

Example:

[ModelFileset]
scan_id = "models"

Defines the location of the trained ML models in a dataset named 'models'.

Database tasks

Clean

Defined in plantdb.task, it is used to clean a scan dataset from the "processing folders".

No luigi parameters (no upstream task).

List of task variables:

• no_confirm: boolean indicating whether a confirmation is required to clean the dataset, False by default, i.e. confirmation required.

Example:

[Clean]
no_confirm = true

Use this to clean without confirmation.

Algorithmic tasks

Colmap task

Defined in plant3dvision.task.colmap, it is used to match scan images and estimate camera poses. It can also be used to compute a sparse and/or dense point cloud.

List of luigi task parameters:

• upstream_task: task upstream of the Colmap task, default is ImagesFilesetExists

List of task variables:

• matcher: images matching method, can be either "exhaustive" (default) or "sequential";
• compute_dense: boolean indicating whether to run the dense Colmap to obtain a dense point cloud, False by default;
• cli_args: dictionary of parameters for Colmap command line prompts;
• align_pcd: boolean indicating whether to align point cloud on calibrated or metadata poses, True by default;
• calibration_scan_id : ID of the calibration scan, requires ???.

Example:

[Colmap]
matcher = "exhaustive"
compute_dense = false
calibration_scan_id = "calib_scan_shortpath"

[Colmap.cli_args.feature_extractor]
"--ImageReader.single_camera" = "1"
"--SiftExtraction.use_gpu" = "1"

[Colmap.cli_args.exhaustive_matcher]
"--SiftMatching.use_gpu" = "1"

[Colmap.cli_args.model_aligner]
"--robust_alignment_max_error" = "10"

Todo

Add a compute_sparse?

Undistorted task

Defined in plant3dvision.task.proc2d, it is used to create undistorted images from pre-computed intrinsic camera parameters.

Masks task

Defined in plant3dvision.task.proc2d, it is used to create binary mask of the plant location within each image.

List of luigi task parameters:

• upstream_task: task upstream of the Masks task, default is Undistorted but can be None.

List of task variables:

• type: method to use to compute masks, choices are: 'linear', 'excess_green', 'vesselness', 'invert';
• parameters: list of scalar parameters, depends on type hence no default values;
• dilation: integer defining the dilation factor to apply when using a binary mask, no default values;
• binarize: boolean indicating whether to binarize the mask, default is True;
• threshold: float used as threshold for binarization step, default is 0.0;

Voxels task

Defined in plant3dvision.task.cl, it is used to reconstruct the 3D structure of the plant from binary or segmented masks.

List of luigi task parameters:

• upstream_task: task upstream of the Colmap task, default is ImagesFilesetExists

List of task variables:

• type: ;

Example:

[Voxels]
upstream_mask = "Segmentation2D"
labels = "[\"background\"]"
voxel_size = 0.3
type = "averaging"
invert = false
use_colmap_poses = true
log = false

Note

Choose 'Segmentation2D' as ` for ML pipeline or'Masks'` for geometrical pipeline.

PointCloud task

Example:

[PointCloud]
level_set_value = 1
background_prior= 0.5
log = false

SegmentedPointCloud task

Example:

[SegmentedPointCloud]
upstream_segmentation = "Segmentation2D"
upstream_task = "PointCloud"
use_colmap_poses = true

Segmentation2D task

Example:

[Segmentation2D]
upstream_task = "Undistorted"
query = "{\"channel\":\"rgb\"}"
model_id = "Resnetdataset_gl_png_896_896_epoch50"
resize = true
binarize = true
dilation = 1
Sx = 896
Sy = 896
epochs = 1
batch = 1
learning_rate = 0.0001
threshold = 0.0035

Visualization task

Example:

[Visualization]
max_image_size = 1500
max_point_cloud_size = 10000
thumbnail_size = 150
pcd_source = "vox2pcd"
mesh_source = "delaunay"

ClusteredMesh task

Example:

[ClusteredMesh]
upstream_task = "SegmentedPointCloud"

AnglesAndInternodes task

Example:

[AnglesAndInternodes]
upstream_task = "ClusteredMesh"

Evaluation tasks

VoxelGroundTruth

Example:

PointCloudGroundTruth

Example:

[PointCloudGroundTruth]
pcd_size = 10000

ClusteredMeshGroundTruth

Example:

PointCloudEvaluation task

Example:

[PointCloudEvaluation]
max_distance = 0.2

PointCloudSegmentationEvaluation task

Segmentation2DEvaluation task