Platform-aware workflows • parallax

The package API stays consistent across platforms, but defaults and typical follow-on tasks differ between aerial and terrestrial data.

Constructors and metadata

aerial <- px_read(
  system.file("extdata", "aerial-demo.csv", package = "parallax"),
  platform = "aerial"
)

scan <- px_read(
  system.file("extdata", "terrestrial-demo.csv", package = "parallax"),
  platform = "terrestrial"
)

aerial$metadata$platform
#> [1] "aerial"
scan$metadata$platform
#> [1] "terrestrial"

The same function, different context

Ground classification runs through one function for both clouds, but the meaning of the result is tied to the acquisition platform.

aerial_ground <- px_classify_ground(aerial)
scan_ground <- px_classify_ground(scan)

c(aerial_ground = aerial_ground$n_ground, terrestrial_ground = scan_ground$n_ground)
#>      aerial_ground terrestrial_ground 
#>                 12                 15

Tree segmentation makes sense for the aerial scene, while plane detection is a better fit for the terrestrial scan.

trees <- px_segment_trees(aerial, min_height = 2.0, crown_threshold = 0.75)
planes <- px_detect_planes(
  px_estimate_normals(scan, k_neighbors = 4L),
  distance_threshold = 0.08,
  min_points = 6L,
  num_iterations = 100L
)

trees$n_segments
#> [1] 3
length(planes)
#> [1] 1

Registration and reconstruction

The installed registration example ships as a source/target pair.

source <- px_read(
  system.file("extdata", "registration-source.csv", package = "parallax"),
  platform = "terrestrial"
)
target <- px_read(
  system.file("extdata", "registration-target.csv", package = "parallax"),
  platform = "terrestrial"
)

icp <- px_icp(source, target, max_correspondence_distance = 2.0)
icp$transformation
#>      [,1] [,2] [,3]  [,4]
#> [1,]    1    0    0 -0.60
#> [2,]    0    1    0  0.25
#> [3,]    0    0    1 -0.15
#> [4,]    0    0    0  1.00

Once a cloud has normals, it can move into reconstruction-oriented functions.

target_with_normals <- px_estimate_normals(target, k_neighbors = 4L)
mesh <- px_poisson_reconstruct(target_with_normals, depth = 4L)

c(vertices = mesh$n_vertices, faces = mesh$n_faces)
#> vertices    faces 
#>        7        6

Takeaway

The package surface is intentionally repetitive: read, filter, segment, register, and reconstruct with the same px_* verbs, and let the cloud class carry the platform context.