Every face the user has ruled on entered the pass as an anchor, and the
scan is exhaustive by design (`dr_face::neighbours`), so a person with
750 confirmed faces cost 750 comparisons against every other face in
the library — and the cost of a library grew with how well it was
named. Most of those comparisons said nothing new: thirty frames from
one afternoon are one point of view, not thirty, and a face that
matches one of them matches the rest.
Each person now enters through at most 100 of their anchored faces
(`dr_face::references`). Eligible are those whose raw embedding is at
least 15 long — one above the gallery floor, since a reference speaks
for someone rather than merely being admitted — with an unmeasured
length admitted as it is everywhere else. From those, the set spanning
the greatest volume is chosen greedily: the longest vector first, then
at each step the face with the largest component orthogonal to the
chosen so far. That is pivoted Gram–Schmidt, and the product of the
residuals it picks is the Gram determinant, so the greedy step is the
exact greedy on the objective. A near-duplicate of a chosen face has
no residual and is passed over; the one profile shot among two hundred
frontal frames is taken early; faces inside the span of the chosen add
no volume and are not taken to fill the cap.
The faces not chosen keep their confirmations and are not touched by
the pass — they stay in the anchor map, so it never releases them —
they are simply not compared. A person none of whose faces is long
enough is still stood for, by their longest, rather than losing their
anchor and having their next face filed as a stranger. Under the cap
nothing changes: every eligible face stands, and the short ones stay
in as the probes they were.
At the reference library's 3,851 confirmations the scan shrinks by
about a fifth; at 15,000 it is a fifth of what it was.
The desktop names where a package may have put libonnxruntime — an
override variable, beside the executable, the package's own library
directory, the Flatpak prefix, the system library directory — and
Android points at the APK's native library directory, which is also
what Qualcomm's DSP loader must be told for the Hexagon skel. Android
starts the engine at the end of the model unpack rather than at launch,
because the probe fingerprints the model files and a first launch has
none until then.
The About panel gains an Inference row beside Graphics, re-read every
two seconds while the probe runs and engines land, and faces.model_id
carries the detector's form: an int8 detector finds a different set of
faces and is a different population (docs/inference.md §7). A
low-memory signal drops every idle session with the GPU caches.
The APK assembly bundles ONNX Runtime and the Qualcomm HTP libraries
from Maven, fetched by tools/fetch-android-runtime.sh with their
published checksums; RUNTIME_DIR=none builds the tract-only APK, which
is a slower app and not a broken one. The desktop packages carry no
runtime yet.
Two probe fixes from the first desktop run: the floor must not be
built with CPU fallback disabled, and a versioned libonnxruntime.so is
a runtime too. On the reference desktop the probe now loads ONNX
Runtime 1.30, measures 30 ms on the CPU provider, and selects TensorRT
at 1.5 ms.
One crate names the runtime, the providers and the devices; dr-face and
dr-segment ask it for a session by role. It hands ort an API table once
per process — from a libonnxruntime it dlopens when the app names a
directory holding one, otherwise from tract — so the Rust build stays
free of C on every target and a package can install the runtime as a
file (docs/inference.md §3).
Sessions live in a registry behind a Model handle that holds the bytes,
not the session: every use refreshes a timestamp and a reaper unloads
whatever sat idle past the decay. A scan that runs the detector on each
image never lets it go idle; a click in the develop view lets the
segmenter go after thirty seconds; a handle used after that reloads,
and reloads on a higher rung if a compiled engine has landed meanwhile.
The probe walks the platform's ladder by building strict sessions and
timing them against the CPU provider, caches the choice against a
fingerprint of the runtime, driver, hardware and models, and compiles
engines for the selected rung in the background, smallest model first.
Nothing in this commit turns the native path on: the apps still run on
tract until they call init with a runtime directory.
The 106 points the eye boxes were cut from, stored beside the reading as
16-bit fixed point over the frame: 424 bytes a face, a seventh of a pixel
on a 6000-pixel frame, where f16 at the same size would have been six.
Derived data like the embedding, kept for the same reason — it cost a
fetch and a model run, and the next per-face pass should run from the
catalog. Shards carry it; a peer's shard from before it is still read.
The register grew both clauses the same day this was built: FR-CULL-8a is
the per-face state the reading is, and FR-CULL-13 is the rule that a
signal is shown and filtered and never writes a judgement. The tags,
faces.md §17 and catalog.md now say which is which; FR-CULL-8a records
what of it is built, and that its third model is under the InsightFace
grant by the same decision as the pair.
SCRFD's eye point places a face, not an eye: on turned and smiling heads
the classifier's window had the eye in a corner, and two model-free ways
of re-centring it — the darkest blob, the most contrasty window — both
lost open eyes (19 → 15 and 19 → 9 of 25). Three landmark models were
then run over the same faces; Face Mesh V2 and InsightFace's 2d106det
tied at 22 of 25 and 2d106det ships, being the cheapest by far and under
the grant the detector and embedder already carry. The eye box is the
tight bounding box of its ten lid points, cut upright from the native
render, which is what the classifier was trained on.
The larger change is that the reading now carries, per eye, the source
pixels across the box and the sharpness of the patch — because the
commonest wrong answer on the reference library was a soft eye read as
closed, and a classifier shown a smear will always say something. An eye
under either floor, or narrower than six tenths of its partner (the far
eye of a turned head, whose contour collapses), is not asked; a face with
no readable eye is a fourth state, Unreadable, that no filter drops. On
twenty native renders the one real blink is caught, the laughing faces
are closed, the profiles are judged on the near eye, and the one thing
left beyond any floor is a face with a pot held over it.
Two MIT classifiers from the same author as the reference pipeline's
whole-body detector: OCEC answers P(open) for one 40×24 eye, SGC
P(sunglasses) for a 48×48 head. Both load in tract once their batch
dimension is pinned by tools/fix-face-model-shapes.sh, like the embedder.
The crops come through the same fitted similarity the aligned face does,
so an eye window is a constant in template units rather than a second
warp, and a tilted head yields an upright eye. Measured on 60 proxies
from the reference library: the eye window plateaus at 22×11, the S
variant beats M and L (which overfit their own domain), and for
sunglasses the aligned face beats a head framing but the higher of the
two catches 11 of 12 pairs against 9 for either alone.
The reading keeps both eyes and the sunglasses number apart, because a
wink averages to the least informative value and a lens of dark glass
draws a confident answer from the eye classifier — over a woman in
sunglasses it read the right eye 0.97 open. Sunglasses take precedence,
and a face behind them is neither open nor a blink.
The embedder's raw output has a length, and the length is a reading of
how recognisable the crop was: a blur, an occlusion or a hard profile
comes out short. Normalising threw it away. A short vector sits near
the middle of the sphere and matches a little of everyone, which is how
one bad crop bridges two people in a grouping pass.
So the length is kept — the store now holds the raw vector, re-normalised
on load, with the length beside it as `faces.quality` — and a face under
MIN_GALLERY_QUALITY (14) is a probe: measured against the gallery and
placed where it fits, but never what another face is measured against.
Two probes are never paired, and a probe is nobody's evidence for a
confidence. The People screen shows the number as "Quality 17.3", dimmed
below the floor.
Faces indexed before this stored unit vectors and have no reading; they
are admitted to the gallery, and schema V14 forgets the run marker of
every image holding one so the next indexing pass measures them. A
peer's unmeasured shard faces are not adopted, or a sync would write
that marker back.
`calibrate.rs` claimed its positive pairs came from user confirmations "then
burst siblings, since FR-CULL-5 already groups bursts", and repeated it beside
the pair floor: "the positives are bootstrapped from bursts and a handful of
early confirmations". Neither is true and neither ever has been. Nothing in
the workspace pushes a burst pair into a `Pairs`; nothing pushes any pair at
all outside this file's own tests. The sentence was written while both halves
of docs/faces.md §8.1 were being planned together, describing a source that
was going to exist, and it has read since as a description of what the code
does.
The distinction matters more here than in most comments, because this file is
the one place in the subsystem allowed to say what a similarity *means*. A
reader who believes the fit is drawing on bursts believes a young library is
gathering positives on its own, which is precisely the opposite of the state
FR-CULL-9 legislates for — a library with no fit, no valid calibration, and a
reference curve it must not present as a measurement of itself. The floor of
200 positive pairs looks arbitrary under the wrong story and obvious under the
right one: confirmations arrive one at a time, from a person.
So the comment now says what is here — a positive is a pair confirmed onto one
person, and there is no second source — and keeps the burst idea where it
belongs, as §8.1's proposal, with the two reasons it is not in the code: this
crate is handed cosines and cannot see a catalog, and the purity of a burst
pair is a thing to measure before it is a thing to trust.
No behaviour changes; the arithmetic is untouched.
Not authored in this session. `cargo fmt --all` reformats every crate, so
running it while working on `dr-segment` picked up five files from the recent
face and library work that had been committed unformatted.
Committed on its own rather than swept into the change that happened to
produce it: the diff is pure whitespace, and mixed into a commit that alters
an algorithm it would be noise in exactly the place someone is trying to read
carefully. `cargo fmt --all -- --check` is a CI gate (tools/ci-local.sh), so
this had to land somewhere regardless.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Implements the FR-CULL-8 written two commits ago. The sweep fetched the
JPEG preview embedded in each RAW and used that one buffer for both
detection and the crop; it now fetches the original, renders it through
the same path export uses, reduces that for the detector, and warps the
crop back out of the native frame.
Three pieces, and each exists for a reason worth stating.
dr_face::Pixels lets the warp sample 8-bit RGBA directly. A 24 MP native
frame is 96 MB as RGBA and 288 MB converted to the f32 RGB align.rs was
written against, and the warp reads about forty thousand pixels out of
it. Converting the whole frame to sample 0.2% of it is NFR-RES-2's
budget spent on a copy, per image, for a whole library. The variant
costs one branch per sample and a test asserts both layouts produce
identical crops.
The detector gets a box-filtered reduction to 1600px, not the native
frame and not a point-sampled one. Averaging rather than sampling
because the detector's job is finding small faces and decimation is
precisely the operation that removes them: at 4x, fifteen of every
sixteen pixels are discarded and a 40px face survives or not depending
on where it falls relative to the sample grid. 1600 rather than 640
leaves the letterbox a mild 2.5x rather than a 9x, and bounds the f32
buffer at 20 MB.
Landmarks come back in the reduction's coordinates and are scaled to
native in one place before any crop pixel is read. This is the failure
mode that would not announce itself -- unscaled landmarks put every crop
near the top-left corner, which yields faces of something else, cleanly
embedded and confidently clustered.
The sweep fetches SWEEP_LANES-wide and renders sequentially. Not a
placeholder for a parallel version: there is one GPU, so concurrent
renders queue on it regardless, and each materialises a native frame.
Overlapping them would multiply the one allocation that threatens the
memory budget while buying parallelism that does not exist. The chunk
drops from 96 to 6 for the same reason -- 96 held 8 MB previews, this
holds whole RAWs.
The stored edit is deliberately not applied, which is where this departs
from export::render_from_library. Face geometry is normalised to the
frame, so indexing a cropped render would record boxes against a frame
that changes whenever the user changes their mind, and every stored box
would quietly become wrong. Orientation is applied: that is a fact about
the file rather than an edit.
examples/face_native.rs renders one file and indexes it both ways, so
the claim behind all of this can be checked against photographs rather
than re-read out of the catalog it came from.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
FR-CULL-8 said detection runs against the thumbnail or proxy tier and
never a full decode, and faces.md §5 said the aligned crop is sampled
from that same proxy. Both are wrong in the same place: they treat
detection and cropping as one resolution problem when they are two, with
opposite answers.
Detection does not care. §4.1 fixes the graph's input at 640x640 and
letterboxes whatever arrives, so a face filling 2% of the frame reaches
the model at 12px whether the buffer handed over is 1024px or 6000px.
Every pixel above the detector's own input is discarded before inference.
The crop cares about nothing else. §5's warp produces the fixed 112x112
ArcFace sees, so source resolution converts directly into whether those
112 pixels were photographed or interpolated. Reading crop_px across the
18,671 faces the proxy-tier implementation stored: 47.3% were upsampled
to reach the embedder, 314 of them by more than 2x, the smallest from 34
source pixels. An upsampled crop does not fail loudly -- it yields a
confident embedding of detail that was never there, and the damage
appears three stages later as clusters that will not separate.
So FR-CULL-8 now specifies four stages with the resolutions named
separately: render native through FR-EXP-9's pipeline, downscale for the
detector, map boxes and landmarks back to native, crop and align from
the native render. The affordability the old rule bought is met instead
by when the pass runs -- background, preempted, resumable -- and the
requirement says plainly what it now costs on a remote library: the
original rather than FR-NC-3's byte range, 412 GB across the reference
library's 19,107 images, so a whole-library pass is a transfer under
FR-NC-6 rather than something that may start on its own.
MIN_CROP_EDGE replaces the MIN_DETECT_EDGE this branch briefly had. Same
number, guarding the quantity that turned out to matter.
faces.md §7b records both measurements, and marks the second as
unexplained rather than dressing it as a finding. Grouped by the buffer
detection ran against, faces per image was 0.078 at 1024 or below and
1.82 at 2048 or better, controlled for file type and size. That gap is
real and reproducible and I cannot account for it, because the letterbox
above says detector input should not matter. M4 is where it gets
settled. The crop measurement does not depend on it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Reverts 53f7cdf and e92d22d. The floor those added sat on
Detector::detect, refusing any buffer under 1025px on the reasoning that
a small buffer finds no faces. That reasoning does not survive §4.1:
the detector letterboxes every input to 640x640, so a face occupying 2%
of the frame presents at 12px to the model whether it is handed a 1024px
buffer or a 6000px one. Detector input is precisely the quantity that
does not matter.
Worse than merely useless, it blocks the design FR-CULL-8 now specifies,
where the detector is deliberately fed a downscale and the crop is taken
from the native render. A guard on detect() rejects exactly that call.
What the measurement actually supports is a floor on the *crop* source,
which is where resolution converts into embedding quality, and which
faces.crop_px already records: 47% of the reference library's faces were
upsampled to reach 112x112. That floor is a separate change against the
native-resolution path and does not belong on the detector.
The 23x faces-per-image gap by source_edge that motivated the original
commit is kept in faces.md §7b, restated as the unexplained observation
it is rather than the causal claim it was written as. V12 stands: those
runs cropped at 1024 whatever detection did, and that is reason enough
to look at them again.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Detection was run on whatever proxy the caller happened to have. A
small one does not fail -- the image is letterboxed into the detector's
640px input at any size -- so it comes back with almost nothing, and the
caller then writes a face_index row saying the photograph was examined.
That row is the damage. Nothing distinguishes it from "examined
properly, no faces in this one", so the image is never looked at again.
The measurement, on the reference library of 23,531 images. Runs against
a 1024-edge proxy: 0.078 faces per image, 90% of them finding nothing at
all. Runs against 2048 or better: 1.82. To rule out the obvious
objection that small proxies just come from small photographs, the same
comparison restricted to DNGs -- 1,592 of them averaging 21 MB against
7,724 averaging 23 MB, so the same kind of file in the same library --
gives 0.078 against 1.82 again. Twenty-three fold, on identical source
material, identical weights, identical options.
So the floor goes in the detector rather than in either sweep, because
both of them, the example tool and any future job handler are equally
entitled to get this wrong, and there is one place that sees every
attempt.
It is 1025, not 1024, and the odd-looking number is the point: 1024 is
exactly ThumbSize::Large, the tier proxies are stored at and the tier
one of the two sweeps was detecting on. A floor that admitted 1024
would admit precisely the population this exists to exclude. Written as
a minimum rather than a maximum so the test at each call site is
`edge < MIN_DETECT_EDGE` with no boundary left to get wrong.
ProxyTooSmall is its own error variant rather than an empty result
because the caller has to tell it apart from a failure: nothing is
wrong with the image or the model, and the answer is to go and find
better pixels, not to retry these ones.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
assign's denominator is the identities the user has ruled on, and it
reads Cluster::person to find them. Master's "Let a name hold a group
together" widened what sets that field: a confirmation, a name, or an
ignore, where before it was a confirmation alone.
The behaviour is right either way — a named person is exactly the
identity a suggestion should be discounted against — but the module note
and faces.md §9.1 both said "a confirmation", which is now too narrow.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Engine::cross is the one place a dot product is computed during
agglomeration — when two groups become adjacent through a third and
their sub-threshold pairs, never summed because they were never
interesting, have to be accounted for. It was calling the portable loop
while the scan beside it had AVX2 or NEON, which on the reference
library was 1,753,514 dot products taking 0.54s.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Seeding the merge heaps was 3.06s of a 5.93s regroup on the reference
18,143-face library — more than half the pass, spent before a single
merge was considered.
Every component scanned the whole pair list looking for the pairs that
were its own: 475 components against 804,499 pairs, 382 million set
lookups to place 804,499 of them. A pair can only ever join two faces of
one component, since that is what a component is, so the union-find that
finds the components can file the pairs at the same time and hand each
agglomeration the list it needs.
The membership set inside agglomerate goes with it — it existed only to
run that filter — and the heap can be sized up front now that the pair
count is known.
Ordering is preserved deliberately: pairs are filed in the order they
arrive, which is the global (i, j) order, so the seeded heap breaks its
ties exactly as before and the merge order is unchanged. Same 2,518
groups holding the same 16,246 faces on the reference library, at 3.5s
rather than 5.9s.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The scan is O(n²) dot products and nothing else, so its speed is the
face subsystem's speed — and it was running at 0.7 flops per cycle.
Two separate faults, both measured over the reference 18,143-face
library on twenty cores. It walked the whole embedding array once per
row, ~336 GB of traffic, where a column tile that fits in L2 is read
once per tile of rows: 4.64s → 2.81s. And the workspace builds for
baseline x86-64 — SSE2, no FMA — into which the portable loop was not
being vectorised at all: 2.81s → 0.86s, 195 GFLOP/s.
So the dot product is now chosen per machine. AVX2 + FMA where
is_x86_feature_detected! finds it; NEON unconditionally on aarch64,
since Advanced SIMD is in that baseline and every Android device the app
builds for has it — with the explicit vfmaq, because LLVM will not fuse
a multiply and an add without being told to. The portable loop stays as
the definition the others are tested against, and
the_fastest_kernel_agrees_with_the_portable_one is the only check the
NEON path gets on a machine that is not aarch64.
Faces::embeddings is one flat buffer rather than a Vec per face: the
pointer chase defeated both the prefetcher and the tiling, and it is
also the layout a GPU pass would want.
Behaviour is unchanged and that is checked rather than asserted — the
same 1,531,969 pairs from all three kernels, and on the real library the
same 2,518 groups holding the same 16,246 faces with the same confidence
distribution. A full regroup there goes from 10.0s to 5.9s; the rest is
the agglomeration, which is a sequential heap walk and is where the next
look should go.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The number beside a suggestion was the mean calibrated probability
between the face and the rest of its group, which measures the wrong
thing twice. It punishes coverage: a person with two hundred faces over
fifteen years is *meant* to have members a given photograph is
orthogonal to, so a correct suggestion onto a well-photographed person
scored low for being well photographed. And it never asked who else the
face might be — a face matching Anna at 0.95 and nobody else, and one
matching Anna at 0.95 and her sister at 0.93, came out identical, when
the second is the only one worth the user's attention.
dr_face::assign answers both, and multiplies them: the mean of the best
ten calibrated matches into the identity (the old mean, capped, which is
what stops coverage counting against it), times that identity's share of
the evidence against every *named* rival.
Only named people compete, and per person rather than per group. Both
halves of that had to be measured on a real 18,000-face library rather
than reasoned about. Normalising across every group made the number
useless — median suggestion 21%, four in five under half — because
clustering leaves one person spread over many groups, so a face competed
against itself; and keying rivals by group left Catherine competing with
Catherine, median 39%. Per named person: median 99.5%.
Rivals are gathered below the merge threshold, down to even odds: a
named person matching at 0.6 will never be merged into but is exactly
the competition to discount for. That would be a second similarity scan,
the expensive half of regrouping a library, so cluster_scored scans once
at the looser floor and hands the merge engine the subset at or above
the threshold — pair for pair what it would have scanned for itself,
held to that by a test.
Leave-one-out over that library's 2,702 confirmations across 54 named
people: 99.33% of faces placed on the right person against the old
mean's 99.15%, and the number shown for the right person moves from a
median of 90.4% to 99.3%. It errs low — 100% correct wherever it states
80% or more — which is the safe direction, and docs/faces.md §9.1 says
plainly that the low bands are not calibrated.
The example that measures it comes too: this is a claim about a
library's numbers, and nobody should have to take it on faith.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
FR-CULL-9 was read as "no fit, no number", so every library without 200
confirmed positive pairs showed "Confidence unavailable" on every
suggestion — which is every library, until enough confirmations exist to
fit one. The confirmations are made on this screen, ranked by the number
it was withholding, so the degraded state was also the permanent one.
There has always been a curve: Calibration::default is the reference
implementation's fitted MBF sigmoid, which is what clustering already
operates at. It is a published operating point, not an invention, and
what the requirement forbids is presenting it *as though it were
measured on this library*. So the percentage is shown, and the screen
says once, above the grid, where the curve came from.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
64 source pixels was too strict: it threw away 70% of everything the detector
finds, and plenty of what it took were faces a person could name.
Lowering it is not a one-line change, because the two floors are coupled. A face
under 112 pixels is *upsampled* to reach the embedder and upsampling invents no
edges, so a small face scores low on sharpness however crisp the original was.
Re-measured over the reference library with `face_index --quality`:
min crop min sharp size cut blur cut kept
32 0.000 44% 0% 56%
32 0.002 44% 3% 53%
32 0.005 44% 8% 48%
32 0.010 44% 16% 40%
32 0.020 44% 27% 30%
64 0.020 70% 7% 23%
Holding the blur floor at 0.020 while dropping the size floor to 32 would have
rejected a further 27% — for being small rather than for being blurred — and
kept only 30%, barely more than the 23% the strict pair kept. Most of the point
of lowering the size floor would have gone straight back out through the other
gate.
0.005 removes 8% of what the size floor leaves, which is the same job 0.020 was
doing at 64 (7%): the large-but-soft face this gate exists for. Together they
now keep 48% of what the detector finds, against 23% before.
The box pre-filter follows down to 24, staying below what the real floor accepts
so it cannot reject a face that would have cleared 32.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The library was storing faces at 52 source pixels and embedding whatever came
back. There was a size floor, but it was 40 pixels on the *bounding box*, and
there was no blur gate at all — so a subject walking through a half-second
exposure detected confidently, aligned cleanly, and produced a perfectly
ordinary-looking 512-vector. Nothing downstream can tell that apart from a real
face, and because blurs resemble each other more than they resemble the people
they were, they cluster together and weld unrelated identities into one group.
Two floors, both measured rather than guessed. `face_index --quality` runs the
detector over real proxies with both gates disabled and prints the distribution;
over 1,503 faces in 600 images of the reference library:
percentile crop px sharpness
1% 16 0.0006
25% 23 0.0025
50% 38 0.0071
75% 76 0.0284
99% 352 0.4282
The median face in a personal library is 38 pixels. Most of what the detector
finds is background: people across a square, a face on a poster, a stranger at
the next table. They are real detections and useless identifications.
**Size, on the crop rather than the box.** "At least 64x64" has to mean the
pixels the *embedder* sees, and the box is not that — the ArcFace template
reaches past it for forehead and chin, so the aligned crop spans roughly 1.3x
the box's shorter edge. The floor is therefore `min_source_px` on the aligned
crop, applied after the warp fixes the scale, and `min_face_px` drops to 48 as
what it always really was: a cheap pre-filter set low enough that it cannot
reject a face the real floor would have kept.
**Sharpness.** Variance of the Laplacian divided by the variance of the luma it
was taken over. The division is the part that matters: raw Laplacian variance
scales with contrast, so a threshold on it would quietly discard every backlit
portrait in the library. The ratio asks how much of the crop's variation is
edges rather than broad gradients, and is invariant to exposure.
What each pair removes, cumulatively, of everything the detector finds:
min crop min sharp size cut blur cut kept
64 0.000 70% 0% 30%
64 0.010 70% 3% 27%
64 0.020 70% 7% 23%
80 0.010 76% 2% 21%
64 and 0.020. The size floor does most of the work, and the blur floor removing
only 7% on top of it is the point rather than a disappointment: at 64 pixels
most faces are already sharp, and what it takes out is the large-but-soft one —
precisely the face that would otherwise contribute a confident, wrong embedding.
The two gates are not independent and the doc comments say so: a face under 112
pixels was upsampled to reach the embedder, and upsampling invents no edges, so
small faces score low on sharpness even when the original was crisp. That is why
`--quality` prints them together.
**This will re-index.** Around 70% of what the current settings store falls below
the new floors — faces between 20 and 40 pixels that nobody could identify. The
People screen gets shorter and every group in it gets better.
66 dr-face tests pass, including that a blurred crop scores below a sharp one,
that halving the contrast does not move the score, and that an upsampled face
scores below the same face at full size.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
0.90 left a third of the reference library ungrouped: 1,213 of 1,813 faces in a
group, and the rest sitting alone in a screen that had nothing to offer for
them.
"Is 0.90 too tight" is not answerable from the number. It is a probability, and
which cosine it lands on depends on the calibration — so the first half of this
is a way to ask the question properly. `face_index --tune` runs the real
clusterer over the real embeddings at ten thresholds and prints what each one
produces. It writes nothing; comparing thresholds by applying them would have
each one pollute the next.
On the reference library:
P cosine groups grouped largest
0.95 0.449 311 62% 51
0.90 0.403 316 67% 51
0.85 0.374 318 70% 57
0.80 0.353 328 74% 69
0.75 0.335 327 77% 69
0.70 0.319 326 79% 81
0.50 0.267 303 85% 90
The count of *groups* is the signal, not the count of grouped faces. Loosening
from 0.95 makes it climb: real people are being assembled out of fragments. It
peaks at 0.80 and then falls — and a falling group count while the grouped faces
keep rising is the shape of over-merging, separate identities being welded
together. That is the FR-CULL-10 failure, and the one the user cannot undo by
hand.
So 0.80: the loosest setting still building people rather than melting them
together. A third more of the library gets grouped than at 0.90, and the largest
group grows by eighteen faces rather than by forty.
The table is one library, and the doc comment says so — `--tune` reruns it on
any other.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Pressing Regroup on a real library did not come back. Clustering 1,813 faces is
the textbook agglomeration — compute every pairwise cosine, then repeatedly scan
all live group pairs, score each with average link, and merge the best — and the
scan is inside the loop. Each merge rescans every surviving pair, and each score
is recomputed from scratch over every cross pair. Some 1.6 million pair scores
per merge, some 700 merges to do.
Three changes, none of which alter the answer.
Only above-threshold pairs can ever matter. An average that reaches the
threshold must have at least one term at or above it, so two groups with no
qualifying pair between them can never merge — not now, and not after any
sequence of merges, since merging only adds terms. The new `neighbours` module
produces exactly that sparse list: 7,875 pairs rather than 1.6 million on the
reference library. It also means the n^2 matrix is never materialised, so memory
goes from O(n^2) to O(edges) — 2.5 GB to a few hundred KB at 25,000 faces.
Merges cannot cross components, so the connected components of that graph are
independent problems: four hundred small agglomerations instead of one large one.
Average link is additive — sum(A u B, C) = sum(A, C) + sum(B, C) — so a merged
group's scores follow by addition. Kept as running (sum, count) per adjacent
pair, a score costs one division instead of a nested loop, and a heap with lazy
invalidation replaces the rescan.
Measured on the reference library: 0.28s, release, for all 1,813 faces.
An exact ANN index was tried and removed, and neighbours.rs records why so it is
not rediscovered as a good idea. IVF with a triangle-inequality bound is exact
and prunes beautifully on synthetic clusters; on real embeddings it prunes
*nothing* — 946 of 946 cell pairs survive. Median pair angle is 88.5 degrees and
the merge threshold is 66.2, so the bound needs cells of radius under ~10
degrees, but two photographs of the same person sit 36-60 degrees apart. No
ball-based partition of a 512-d near-orthogonal space can be tight enough. So
the scan stayed exhaustive and got an unrolled dot product and its blocks spread
across cores instead.
Correctness is held by keeping the old implementation as an oracle: three tests
run both engines over the same population — plain, under co-occurrence and
anchor constraints, and with a size-weighted calibration — and assert the
clusters are identical. Determinism is asserted at a size where the threaded
path is in play.
62 tests pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The merge of the SCRFD/MobileFaceNet work brought 69 rustfmt diffs across
dr-catalog, dr-face and dr-ui with it, so `cargo fmt --all -- --check` fails
on master and the Desktop job stops at its Format step — before clippy, the
tests or the release build have run at all. That makes the whole desktop
half of CI blind: a real compile error behind this would look exactly the
same from the outside. There was nothing behind it, as it turns out — with
the formatting fixed, clippy, the test suite and the release build all pass.
Every .rs hunk is `cargo fmt --all` on the pinned 1.92.0 toolchain, not a
hand edit, but it is worth being precise about what that moved, because it
is more than whitespace. Besides reflowing signatures and call chains,
rustfmt reordered the `pub mod` and `pub use` items in dr-face/src/lib.rs so
the `#[cfg(feature = "inference")]` entries sort in place, added the trailing
semicolon inside `let ... else { return }` bodies in identity_ui.rs, wrapped
a bare closure body in braces in cluster.rs, adjusted trailing commas, and
dropped a stray blank line at the end of identity_ui.rs. All of it is
semantically inert; none of it changes behaviour.
docs/traceability.md rides along because it has to. The matrix records each
TRACES tag by line number, and reflowing develop.rs, lib.rs, faces.rs,
identity.rs and identity_ui.rs moved them — FR-CAT-8, FR-CAT-9, FR-CULL-10,
FR-DEV-3, FR-DEV-3a and FR-DEV-3c all shift by a line or two. The matrix was
verified up to date on d777f7f before this commit, so this is drift these
formatting changes introduced, not pre-existing staleness being swept up.
Leaving it for a follow-up commit would hand traceability-check.yml a
failure caused entirely by a whitespace change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Faces are found on the thumbnail, which is cached the right way up --
the grid would lie on its side otherwise. Segmentation runs on a proxy
rendered through a neutral edit graph, which carries no orientation and
is therefore in sensor order. For anything shot in portrait the two
differ by a quarter turn, so a face and the person containing it were
being compared in spaces 90 degrees apart: no match, or worse, a match
against somebody else's region.
The transform goes on the face rather than on the proxy. Instance masks
are defined in the proxy's space and sampled long afterwards, so turning
that space would be a far larger change than naming a region warrants.
Also two things the first screenshot of the running app showed that no
test would have:
110 of 23,528 displayed as "0%", which reads as the feature having done
nothing. One decimal below ten percent, and a floor so real progress
never shows as none.
The rail picked some near-black covers, because the largest face in a
group is often the nearest one in a badly lit frame and a black square
beside a name identifies nobody. It now cuts the best few and takes the
first legible one, falling back to the largest when a person's every
photograph is dark -- which happens, and showing it beats showing
nothing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The segmenter knows it found a person; the face index knows which person.
Joining them turns "person" in the mask list into "Anna", which is the
difference between a vocabulary of eighty COCO classes and one that
includes the user's family. Selecting a subject in a group photograph
stops being a guessing game between three identical rows.
Containment, not IoU. A face is a small part of the person it belongs to,
so a correct pairing has an IoU near zero and anything IoU-based would
reject every true match.
Confirmed names only. A suggestion is the system's guess, and printing a
guessed name onto a mask region would launder it into a fact.
Writing the tests corrected the design once: a tight head-and-shoulders
portrait, where the face fills most of the person box, is the case where
naming is most certain, not least. An earlier guard rejected exactly that
and has been removed, with the reasoning left as a test because it is
easy to get backwards a second time.
The names hang on the develop session, set when the image opens because
that is the one moment the catalog and the image id are both in reach.
Every segmentation run afterwards picks them up for free, and a library
with no face indexing behaves exactly as it did before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Wires dr-face to dr-catalog: a background sweep that reads the proxy the
grid already built, detects, aligns, embeds and stores, then a clustering
pass that turns those embeddings into suggested people.
Detection runs on the Large thumbnail tier and nowhere else. That is what
makes the feature affordable -- a browsed library has already paid for
its proxies, so face indexing adds no RAW decode that was not already
happening -- and it is why an image whose proxy is missing is skipped
rather than fetched: requesting one here would put face indexing on the
network path FR-CULL-8 keeps it off.
The sweep keeps no cursor. It asks the catalog what is missing, so it
resumes after process death with no repeated work beyond the in-flight
image, and cancelling is dropping the receiver.
recluster writes only the suggested half. Confirmed faces go in as
anchors and come back untouched, and a cluster of one stays nameless --
naming every stray face would fill the People view with noise the user
then has to dismiss.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
FR-CULL-9 forbids thresholding a bare cosine anywhere in the subsystem,
so calibrate fits P(same person) per library and reports whether the fit
is trustworthy. Two details carry most of the weight.
The fit runs against a 200-bin histogram rather than a pair list: a
25,000-face library has ~3e8 pairs and no gradient descent is running
over that. And a fresh library has no valid calibration, because the
positives have to come from user confirmations or burst siblings --
bootstrapping them from high cosine would fit the calibration to the
belief it was supposed to test.
Clustering defends against the over-merging FR-CULL-10 warns about with
constraints rather than a better threshold: two faces in one photograph
never merge, and two groups confirmed as different people never merge.
Average link rather than single link, so one strong edge cannot weld two
families together.
Calibration is defined once, in dr-face, and dr-catalog re-exports it.
Two implementations of one probability model is exactly how a number
comes to mean the wrong thing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Ports the pipeline from the C++ reference in ../scene-actor-extraction
(MIT, same author). End to end on real portraits it separates identities
the way the reference's fitted calibration says it should: 0.596 between
distinct photographs of one person, 0.05 between different people, either
side of MBF's 0.267 boundary.
Three things are structural rather than incidental:
Aligned112 can only be built by align::warp, so Embedder::embed cannot be
handed an unaligned bounding-box crop. That mistake yields 512 plausible
unit-norm numbers and no error, so the type system refuses it instead.
Embedding carries its ModelId and cosine() returns None across models,
because a cross-model similarity is the one mistake that produces
plausible garbage rather than a failure.
The model-free half -- alignment, embedding arithmetic, f16 storage --
sits outside the inference feature and is covered by 11 tests that need
no weights on the machine.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Neither InsightFace export parses as shipped -- SCRFD fails at its input
node, ArcFace at the first Conv -- which is the same wall dr-segment hit
on YOLO's dynamic export. Both load cleanly with the input dims frozen,
so the pure-Rust runtime holds for the face pipeline too.
tools/fix-face-model-shapes.sh does the freezing, and exists so the
artefact is reproducible rather than a binary someone once produced. It
takes two forms because the two graphs need different ones: ArcFace's
batch is a named dim_param, SCRFD's H and W are dynamic but unnamed.
Also notes YuNet loading with no intervention, which matters for the
licence question in faces.md 2.3.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
dtourolle
