Lazy multiprocess target extraction

[davidtvs]

When using longer datasets or larger image resolutions, kalibr uses a large amount of memory during corner extraction.

That's because all inputs needed for the corner extraction (including all images in the dataset) are duplicated numProcesses times when each process forks.

This PR simplifies the code and uses lazy multiprocessing to feed inputs to the process pool only when needed (one-by-one).
See a comparison below of the memory consumption before and after this PR with a dataset of 1920x1080 images:

Peak memory usage is reduced by more than 50%.

CPU usage is more or less the same - seems less "spiky" now:

I've also ran the code in the PR on the sample dataset to make sure there's no unintended degradation:

Command:

rosrun kalibr kalibr_calibrate_cameras \
 	--target april_6x6.yaml \
 	--models pinhole-radtan pinhole-radtan \
 	--topics /cam0/image_raw /cam1/image_raw \
 	--bag cam_april.bag \
 	--bag-freq 10.0

Camera chain before this PR:

cam0:
  cam_overlaps: [1]
  camera_model: pinhole
  distortion_coeffs: [-0.2894411946323578, 0.07962117717814395, 0.00019362566733012107, 2.0633489114325153e-05]
  distortion_model: radtan
  intrinsics: [459.16620162518655, 457.8283425910929, 366.56421591367956, 248.93486745674633]
  resolution: [752, 480]
  rostopic: /cam0/image_raw
cam1:
  T_cn_cnm1:
  - [0.9999975833780851, 0.0021876991183658436, 0.00021727987704355133, -0.10996992818618531]
  - [-0.0021905046383791527, 0.9999012408498518, 0.01388201124994388, 0.0003967599845687264]
  - [-0.0001868887548949044, -0.013882453654949846, 0.9999036166316727, -0.0006390066744876479]
  - [0.0, 0.0, 0.0, 1.0]
  cam_overlaps: [0]
  camera_model: pinhole
  distortion_coeffs: [-0.28702698964616846, 0.0774641314753513, -9.0868908124502e-05, 1.191942692999466e-05]
  distortion_model: radtan
  intrinsics: [457.8635851933246, 456.4589896582416, 379.35599660413214, 255.87526793697964]
  resolution: [752, 480]
  rostopic: /cam1/image_raw

Camera chain after this PR:

cam0:
  cam_overlaps: [1]
  camera_model: pinhole
  distortion_coeffs: [-0.28646742048247464, 0.07664440221166424, 0.00037989530732547436, -4.092622872711382e-06]
  distortion_model: radtan
  intrinsics: [459.2480732747205, 457.9630133876875, 366.85529050387214, 248.14762195518594]
  resolution: [752, 480]
  rostopic: /cam0/image_raw
cam1:
  T_cn_cnm1:
  - [0.9999972253522178, 0.0022194793620354293, 0.0007894296777357621, -0.10995275408589618]
  - [-0.002230016923524006, 0.9999049229804217, 0.013607792768362957, 0.0004232690628729674]
  - [-0.0007591524059026039, -0.01360951545307229, 0.999907098072989, -0.0005893580755146163]
  - [0.0, 0.0, 0.0, 1.0]
  cam_overlaps: [0]
  camera_model: pinhole
  distortion_coeffs: [-0.2861004711786593, 0.07657158365440236, 5.9945048063776784e-05, -7.586504253593691e-06]
  distortion_model: radtan
  intrinsics: [458.1502957674771, 456.7543594159104, 379.3836585373581, 255.19745143306127]
  resolution: [752, 480]
  rostopic: /cam1/image_raw

There's small difference but I think that's just the run-to-run variance.

Maybe this addresses #22, hard to say since that issue is very old.