A survey is more than a count: what tells you about a population is
which animals are in it. This release adds the demographic step,
following "When One Modality Is Not Enough: Multimodal Sex and
Life-Stage Classification of Red Deer from Aerial RGB-Thermal Video".
Tracked animals are classified by species and sex, juveniles are
flagged by size, and the two cameras are used together rather than
separately - because they fail in opposite conditions. In colour a
deer under canopy blends into the ground; in thermal it is an
unmistakable blob but the fur colour is gone and antlers show only
while they are still growing and warm. Which sensor carries the sex
cue therefore changes with the season, and reading both at once is
what stays reliable across the year.
    The classifiers read features from Meta's DINOv3, which is a gated
model on Hugging Face: request access once and enter a read token in
the new Classification tab. Nothing else in the plugin changes if you
do not use it.
    Added:
    - Segmentation tool: Source selector - the extracted frames, or
the P6 orthomosaic. The mosaic is cut into overlapping tiles, each
tile is segmented, and a prompt's masks are merged back into one map,
so a tree canopy seen in ten frames is one polygon with one id instead
of ten duplicates. The mosaic's GeoTIFF transform geo-references the
masks directly; no DEM step. Thermal mosaics are stretched to 8-bit
for the model.
    - Segmentation tool (toolbar): SAM3 moved out of the Processing
tab into a window of its own, because prompting is interactive. It
shows the extracted frames of either camera, takes text prompts or
clicked points (left = on the object, right = beside it, several
objects per frame), and segments the current frame, a frame range, or
tracks the objects across the range as one clip with the video
tracker. Three backends: the Roboflow API (text on images),
facebook/sam3 through transformers (everything, CPU or GPU), and
Meta's official sam3 package, which is the only way to run SAM 3.1 -
Meta's March 2026 update with object multiplexing for ~7x faster
many-object video tracking, published on Hugging Face as a bare
checkpoint without a transformers integration. Geo-referencing, the
QGIS layers and the GeoJSON export of the masks live in the window
too; results merge per prompt, so a re-run of "deer" leaves "boar"
alone. The SAM3 configuration tab is gone - backend, model and prompts
are remembered per machine, not per flight.
    - ALFS-PY 3.0: the PyTorch renderer is no longer a separate
package. One framework now carries three render engines (ModernGL,
PyTorch, Vulkan) and two ray casters (Embree, Warp), and the
Dependency Manager picks them from two dropdowns instead of a PyTorch
checkbox. The choice is saved in the QGIS project and exported to the
renderer on every step that renders, so a project renders the same way
whoever opens it - and Install fetches exactly what that choice needs.
A machine still carrying the old AlfsTorch package is told, and
Install removes it: it provided the same import name and would
otherwise silently override the selection.
    - Species classification has a published default model
(cpraschl/bambi-species-classification on Hugging Face): red deer, roe
deer and wild boar, downloaded on first use like the occlusion and sex
classifiers. Its class names map onto the project's species without
configuration. The default crop padding is 0.25 - a square at 1.5x the
longer box side, which is what all the published classifiers were
trained on; a project embedded at another padding simply embeds again
into a new run.
    - Classification (Processing tab): occlusion, species, sex and age
classifiers on tracked animals, each its own step so they can be run
separately or repeated one at a time, plus a size-based age estimate
where no classifier is configured. Species and sex are decided per
animal by a vote across its frames, which is what makes a noisy
per-frame call safe - an antler resolves only from some angles, so
many frames of a true male look female and the majority still recovers
him. The margin behind every call is kept, so a borderline animal can
be reviewed, and the vote can be repeated at a different quorum
without re-running anything.
    - C1 Match RGB and Thermal Tracks: works out which thermal track
and which RGB track are the same animal, by registering the two views
onto each other and comparing where the boxes sit. An animal both
cameras saw is a confirmed animal; a track only one camera saw is
either an animal the other sensor cannot make out, or noise. The run
log reports the confirmed count, and when nothing matches it names
which gate rejected everything and how far the closest candidate was -
so "there were no animals" is distinguishable from "the gate is wrong
for this resolution". Matched pairs can be added to QGIS as a line
layer.
    - C2 Compute DINOv3 Embeddings: describes every animal's crop
once, so all three classifiers reuse the same features. Vectors are
written beside the frames, one file per frame, and are reusable
outside the plugin. A re-run embeds only what is missing, so an
interrupted run resumes rather than starting again, and changing the
crop settings starts a new set without discarding the old one.
    - C6 Age Classification: a juvenile cannot be told from an adult
female by appearance at survey resolution - which is why the sex
classifier's second class is female/juvenile - so size settles it. An
animal is called a juvenile only if it sits far below its cohort and
has a clear gap to the next animal up; in any herd someone is
smallest, and that alone is not evidence. Sizes are compared only
within one flight, because how tightly boxes fit varies between
recordings. Needs no models.
    - Classification configuration tab: the Hugging Face token (kept
in the QGIS settings, not the project file, because a project gets
shared), a "Check access" button that answers before a long run
starts, the classifier table, crop and voting settings, and the
cross-modal matching gates.
    - "Download models" in the Classification tab fetches every
classifier set to Default, so the class mapping can be set up before
anything is run - asking someone to run a classifier once before they
can configure it is backwards. The classifier files are a few
megabytes each; the DINOv3 model is downloaded separately on the first
embedding run.
    - Class mapping per classifier: connects the classes a model
returns to the values used in this project. The mapping follows the
class order rather than the names, since a model returns positions -
so renaming a label never re-points it, and a model that does not name
its classes at all still works. Classes are read from the model where
possible, probed for their count where the model carries no names, and
defined by hand where the model is not available yet.
    - Sex and life-stage classifiers per species: the cue is
species-specific - antlers, for red deer - so a model fitted on one
species says nothing useful about another. A species with no sex
classifier is simply not sexed, which is the honest answer rather than
a guess. Life stage offers a third choice per species, "Size-based",
which is the box-area measurement rather than a model, and is the
default because no life-stage model has been published yet. Which of
the two decides a species is one decision in one place rather than a
model choice plus a switch elsewhere.
    - Life stage is measured on the geo-referenced boxes where
geo-referencing has run, which makes the areas metric; otherwise on
the camera-frame boxes, which still work because the comparison never
leaves the flight. Which was used is recorded with every verdict, so
metric and pixel figures are never mixed. Animals a classifier already
called still count towards the cohort statistics - excluding them
would shift everyone else's score - they simply do not take a verdict
from it.
    - Occlusion is optional. It selects which frames carry usable
evidence, but species and sex run either way: over the frames the
occlusion classifier called clear, or over occlusion values you
annotated by hand, or over every frame. The run log always says which
of the three it used, so voting over occluded frames is never silent.
    - Results are written onto the animals themselves - species, sex,
life stage and per-frame occlusion - which is what makes them visible
to the exports, the map layers, the survey analytics and the labelling
tool. Optional, repeatable, and it never changes a track you annotated
by hand or a species the detector itself identified.
    - Both perspective and orthorectified crops are supported, with
the classifier variant following the imagery so a model always sees
the kind of image it was trained on.
    - Dependency Manager: a Classification group installing
transformers and huggingface-hub. The same packages serve local SAM3
segmentation, which needs transformers 5.0 or newer.
    - Export: a GeoJSON exporter for the SAM3 segmentation masks,
writing their outlines as polygons in WGS84 longitude/latitude, as
GeoJSON requires. Masks that were never geo-referenced are skipped and
reported rather than written in pixel coordinates.
    - S1 Run SAM3 Segmentation can run on your own machine. "Run SAM3
locally" in the SAM3 Segmentation tab loads Meta's facebook/sam3
through transformers instead of sending every frame to Roboflow, so no
Roboflow key is needed. The model is gated on Hugging Face exactly
like DINOv3: request access once, and the token from the
Classification tab downloads it into the shared model cache on the
first run. A "Check access" button answers before the frames are
loaded, the model field accepts a fine-tuned repository or a pinned
revision, and the device follows the Classification tab. The masks are
written in the same structure as the Roboflow results, so
geo-referencing and the exports do not care which backend produced
them.
    - Track details in the inspector: the Feature Viewer now lists
what the project knows about the track it is showing - the same facts
as its row in the track inventory: species, sex and age with their
votes, box and occlusion counts, confidences, frames and times,
positions, movement, the matched track on the other camera and the
labelling-tool annotation - with the inventory's Approved checkmark
beside it. A clicked detection shows the track it belongs to. Ticking
Approved in the viewer or in the report is the same verdict, and an
open report follows.
    - Reviewing in the inspector: a wrong result can be deleted from
the project in the Feature Viewer. "Delete detection" removes the
highlighted box on the current frame - the detection, its ground
position, its place in the track, its classifier results and its
match; the track keeps its other boxes, and a track whose last box
goes is removed with it. "Delete track" removes the track with every
one of its detections; a track drawn in the labelling tool takes its
annotation with it. Both ask for confirmation and cannot be undone.
The map follows: the box leaves the detection layers, a track that
lost a box is redrawn from the ones it still has, and a deleted
track's layers, group, matched-pair line and inventory point are
removed - so no orphan boxes stay behind. An open track inventory
report drops or refreshes the row, and the written inventory, density,
distance-sampling and population results are marked stale, since their
counts no longer hold. Boxes of a label track are not deleted one by
one here; they are defined by its key frames and are edited in the
Labelling Tool.
    Changed:
    - The GeoJSON export is now named "GeoJSON (animals)" alongside
the new "GeoJSON (segmentations)". The old name described how the file
was made rather than what is in it, which is no help when choosing
between two of them.
    - New projects seed the occlusion vocabulary as clear/occluded,
matching what the occlusion classifier reports, so predictions and
hand annotations share one vocabulary. Projects created earlier keep
the values they have - enum ids are append-only and are never
renumbered - and a classifier is pointed at them through its class
mapping. Migrating a 5.x flight appends whatever levels it carries
rather than dropping them.
    - The Processing tab is grouped into Detection and Tracking,
Classification, and Segmentation. They are not one sequence:
classification works on animals the first section found, and
segmentation neither needs that section nor feeds into it. Steps are
prefixed A, C and S accordingly, so SAM3 segmentation is now S1.
    Fixed:
    - GeoJSON exports wrote the project CRS - usually UTM metres -
into a format that has exactly one coordinate system, WGS84
longitude/latitude. The crs member that named the projection was
dropped from the GeoJSON specification years ago and readers ignore
it, so every animal landed at a longitude of several hundred thousand
degrees. Both GeoJSON exports now project to WGS84 and write no crs
member; an export without a known project CRS is refused rather than
written wrong.
    - Camtrap DP media.csv left the timestamp column empty. The poses
store each frame's capture time as an ISO-8601 string with the drone
clock's timezone, which the exporter took for epoch seconds and
silently failed to parse. Both forms are read now, and the timestamps
are written at second precision with their offset; this also fills
deploymentStart and deploymentEnd in deployments.csv and eventDate in
the Darwin Core Archive, which went through the same code.

None