Drift correction & particle segmentation — plus the four defects that made Segment unusable on real data - #117

Open

CSSFrancis wants to merge 38 commits into

mainfrom

feat/drift-particles

Commits on Aug 2, 2026

feat(drift,particles): rigid drift solve + particle container and measurement
CSSFrancis
committed
a8dbfaf
feat(data,drift): synthetic particle-movie fixture, and two solver bugs it found
CSSFrancis
committed
16ef8c8
chore(scripts): one-command verification and benchmark for drift + particles
CSSFrancis
committed
008406d
feat(particles): frame-to-frame linker and the birth/death/merge/split stream
CSSFrancis
committed
08effdd
docs(plan): two overlay defects found by rendering real linker output
CSSFrancis
committed
5358fdf
feat(particles): torch feature stack and scribble classifier
CSSFrancis
committed
b8dfc28
docs(plan): min_size is where the sensitivity trade lands, and min_size=0 is a footgun
CSSFrancis
committed
53bce6f
feat(particles): the particle TREE, plus the requires_particles gate
CSSFrancis
committed
04f12f0
feat(renderer): generic DataTable and a resizable bottom dock
CSSFrancis
committed
2e484c4
feat(actions): seg_ and drift_ staged wizards, and two fixes they exposed
CSSFrancis
committed
aff90c0
feat(renderer): Segment and Drift carets, plus two backend bugs they exposed
CSSFrancis
committed
2f70355
feat(drift): correlate on an alignment ROI
CSSFrancis
committed
4ac6604
docs(plan): the caret shows ONE control; everything else is Advanced
CSSFrancis
committed
4d988fe
feat(particles): live overlay, editing, and the three navigator lanes
CSSFrancis
committed
278824d
fix(signal_tree): close() leaked the new wizards and pinned the source movie
CSSFrancis
committed
171e4ea
refactor(renderer): the Segment caret is one slider and a button
CSSFrancis
committed
a194dd8
feat(drift): ROI discovery, dy/dx as its own plot, and a caret of two toggles
CSSFrancis
committed
500dad9
perf(seg): bound the preview by AREA -- 8.4 s to 0.4 s on a 4k frame
CSSFrancis
committed
660cfe7
fix(seg): actually create the brush -- scribbling was impossible
CSSFrancis
committed
e6818d8
fix(seg): the strip's class and eraser never reached the brush
CSSFrancis
committed
197fe90
perf(particles): decimate the split geometry -- 4k segmentation 7.1s to 2.8s
CSSFrancis
committed
8412dd2
fix(seg): the caret's params decide the class, not the JS widget
CSSFrancis
committed
0f5e548
wip(particles): restore the in-situ working set
CSSFrancis
committed
7d18963
feat(drift): non-rigid correction -- scan-knot and dense field, one solver (A2-A5)
CSSFrancis
committed
6a4d72a
feat(drift): wire non-rigid into the caret, and measure it at 4096^2 x 300
CSSFrancis
committed
aff8d6d
perf(drift): apply the non-rigid field on the GPU -- 392 -> 48 ms at 4096^2
CSSFrancis
committed
8e597ae
fix(drift): the non-rigid tab was still locked in the renderer -- feature unreachable
CSSFrancis
committed
e89f10a
fix(seg): the drift-padded border segments as a giant particle when inverting
CSSFrancis
committed
bc39fe8
perf(dask): the navigator fill submitted ONE task at a time -- 970 round trips -> 2
CSSFrancis
committed
6d4e6dc
fix(nav): the threaded navigator fill never noticed the cluster arriving
CSSFrancis
committed
cc93819
fix(seg): the preview-window box only ever appeared under Classical
CSSFrancis
committed
664e7f7
feat(seg): one Confidence slider that means "fewer particles" on every engine
CSSFrancis
committed
3a343fb
perf(seg): draw ONE raster mask above 100 particles instead of N polygons
CSSFrancis
committed
87a0d75
perf(seg): watershed per COMPONENT, not per frame -- 852 MB/frame was pausing workers
CSSFrancis
committed
175e3e7
fix(seg): the overlay's raster path could never run on a tiled frame
CSSFrancis
committed
514f427
fix(seg): the caret threw on first render, and three controls lied
CSSFrancis
committed
c835fa0
test(seg): a fixture that can actually FAIL, and the spec that pins it
CSSFrancis
committed
fb54447
docs(seg): the measured numbers, and the two traps that hid all of this
CSSFrancis
committed
3ea69d2