Detection and tracking

Assume volume is a reconstructed Float32 array on any backend.

Detect one frame

threshold = 30 / 255
binary = volume .<= threshold
binary = dilate(dilate(binary))

boxes = particle_bounding_boxes(binary)
particles = particle_coor_diams(boxes, volume)
dictsave("particles.json", particles)

dilate remains on CPU, Metal, or CUDA. Connected components stage one slice at a time to the host, and coordinate metrics stage one bounding subvolume at a time. This bounds memory and keeps the public workflow identical.

Choose particle_bounding_boxes_3d if a component must touch the immediately previous z slice. Use the default non-strict function for the historical holographic ghost-suppression behaviour.

Output schema

Dict(
  UUID(...) => Float32[x, y, z],
  # or Float32[x, y, z, equivalent_diameter]
)

Coordinates are one-based pixel/slice coordinates. Multiply x and y by pixel pitch and z by slice spacing, then add your reconstructed-volume origin to obtain physical coordinates.

Link successive frames

frames = dictload.(["000001.json", "000002.json", "000003.json"])
graphs = [labonte(a, b) for (a, b) in zip(frames[1:end-1], frames[2:end])]
paths = enum_edge(first(graphs))
for graph in Iterators.drop(graphs, 1)
    append_path!(paths, graph)
end
full = gen_fulldict(frames)

Tune max_distance in labonte to the maximum plausible frame-to-frame motion in coordinate units. dim3weight scales the less precise depth axis. The function does not mutate the input dictionaries and rejects UUID reuse across frames.

Plotting is optional

using Plots # activates ParticleHolographyPlotsExt
particleplot(frames[1]; scaling=(10.0, 10.0, -100.0))
trajectoryplot(paths, full; scaling=(10.0, 10.0, -100.0))

The core package does not load a display stack. On a compute server, omit Plots.jl and save particle dictionaries for later visualisation.