Product name decided the pipeline: Mini worked live but replay refused it. A bundle descriptor now declares profile_id, product_family and processing_backend, with computational and visualization assets kept apart, and one resolver serves live and recorded input. Legacy bundles load through a documented adapter that never guesses a backend from a folder name.
Pause now holds the pipeline without tearing it down: CAPTURE_ONLY keeps the camera
streaming so a live sensor stays converged while the flow and HHD stages stop, and BOTH stops
the read as well. Resume rebases the tracker, because a held reference frame would otherwise
report the hold itself as deformation.
Verified with tests/test_pipeline_pause.py: 10 passed across three consecutive runs.
The two canonical examples taught the advanced path: they sequenced
get_img() -> compute_deformation() -> compute_contact() -> read_info() by hand
and unpacked a 13-element tuple positionally, which is what led developers to
reach for the compute backend, the capture queue and the stitch mode. The
managed alternative already existed and was documented nowhere.
Library (add-only, nothing removed):
- FrameSnapshot.image_size_raw carries the capture resolution, so a caller can
compare it with the processed frame instead of reaching into Sensor.
- StreamPipeline.reset() resets the tracker and force filters *safely*. It is
queued and applied by the flow stage before its next frame, never inline:
Sensor.reset rebases initial_img on the current Sensor.img, which the flow
stage is reading concurrently, so a caller-thread reset would capture a
half-written reference frame. Idempotent.
- TactileProcessor.pipeline(produce) and FingerRuntime.pipeline() build a
pipeline without the caller naming a backend concept; the latter is
`with runtime.pipeline(contact=True) as pipe`.
Examples:
- examples/basic_pipeline.py and examples/finger_pipeline.py rewritten onto
StreamPipeline + FrameSnapshot: one object, one call per frame, no tuple, and
no compute backend, queue or stitch setting in sight.
- examples/archive/basic_pipeline_manual.py and _finger_pipeline_manual.py keep
the previous versions byte-for-byte apart from a supersede header, so code
copied from them keeps working. The finger header also records the archived
--show defect (store_true with default=True, so it could never be disabled).
- examples/_display.py holds the shared display helpers, so an example no longer
hand-rolls depth normalisation and colormap application.
- examples/README.md leads with the recommended pair and documents the manual
form as still-supported, and states that examples/gui/ is moving to
orisys-studio.
Tests: three added to tests/test_pipeline.py, covering the reset queue (it must
not run on the caller's thread and must land before the first compute_flow),
image_size_raw, and the None case when a producer reports no size. Verified the
reset test fails when reset() is made inline, so it discriminates.
Verified end-to-end against real media: basic_pipeline on data/sample/s1_480p.mp4
(box) reaches ~113 FPS and exits 0; finger_pipeline on data/dd02-0703/1.mp4
(finger) reaches ~140 FPS with mask coverage 0.41 and exits 0. Both archived
manual versions still run. tests/test_pipeline.py: 20 passed. Full suite:
342 passed, 3 failed - all three failures reproduce on the pristine tree
(import precedence, longrun timing, and a stale ruff version literal in
docs/code-review.md) and are unrelated to this change.
ykcheungab
zhangyijian