PyramidBuilder as cuboids app

[JohnTigue]

• Want cuboid api to enable stepfunction code agains cuboids, and it scales as needed
• Under the hood it's Lambda distributed via chunkflow/SQS
• Leverage Countless' fast mode
  • i.e. the one that memoizes (2,2,2) => (1,1,1)
  • (2,2,2):(1,1,1) :: 8:1
  • countless cares about categorical data. Irrelevant here; want memoization.
• The Lambdas are the slaves building the pyramid.
• Chuckflow already implement countless. See also #16

Point is: FFN and bossDB are not storing cuboids as cubes and this slows down the down/up sample pyramid building. They were storing physical cubes of anisotropic voxels; perhaps store digital cubes is more scalable. So cubes via Lamba might be very fast and scalable.

[JohnTigue]

Granted, COUNTLESS is concerned with segmentation labels, not pyramids of raw cuboids from microscopy imaging. Nonetheless, it's relevant. First build pyramids for raw images and later pyramids of segmentation labels. Even the output of brightfield reconstructors will have multiple segments.

COUNTLESS papers

• Original 2D
• Extension to 3D

Note that the latter states:

While COUNTLESS 2D can outperform naive algorithms, using COUNTLESS 3D for high throughput applications is not advisable as it is much slower than the standard approach. However, in Python, it can be implemented without resorting to extra compilation steps and still provides reasonable performance.

• [ ] What about numba?