SOL Vision object registry

Deployed October 10, 2026. Local viewer:
<http://127.0.0.1:8888/api/dashboard/object-registry/view>
or the workstation's current LAN address on port 8888. Public clients receive
403 for the viewer, frame input, JSON, and evidence routes.

What is running

sol-object-registry.service samples both existing camera managers every ten
seconds. It never opens a V4L2 device or substitutes a camera. An atomic local
frame route supplies the original upright image, capture ID, time, freshness,
and screen-candidate geometry. Stale frames, warm-up placeholder images, and
duplicate capture IDs cannot create sightings or missing-object events.

The worker reuses the installed YOLO11n ONNX model through OpenVINO on CPU,
with two inference threads and a 0.35 reporting threshold. Eighty ordinary
object categories are available. Results are provisional model labels,
not accepted identities. The spider retains its user-confirmed visual
references, two recorded views, and rotation/scale reacquisition. Its name
remains unknown spider entity, and only its face is overlaid.

The registry adds no household alerts, person identities, or crossing counts.
It does not draw generic detector guesses on the public camera feed.

Records and meanings

- SQLite authority: /mnt/sol-data/sol-stack/object-registry/objects.sqlite3.
- Named objects have durable obj- identifiers. The imported spider is
obj-45c73f1de776426e; its history begins at registry installation, not at the
creation times of older reference images.
- Generic detections have track- episode IDs. Association requires the same
camera, category, detector, recent observation, and overlapping box, with
one-to-one assignment. It is spatial continuity, not visual re-identification.
After an episode is lost, a later generic detection gets a new episode.
- tentative: one generic detection. observed: a named-reference match or
repeated generic evidence. Neither confirms a model's semantic class.
- not_seen: three fresh samples did not match; not proof of departure.
- unknown: camera or detector observations unavailable; never counted as
absence. The viewer also expires a stopped worker's current states.
- Events include registration, first observation, repeated detection,
reacquisition, unavailability, not-seen, and sampled sightings. Repeated
sightings are written at most once per minute per object.
- Evidence crops are local, mode 0600, and SHA-256 recorded. The generated
evidence budget is 512 MiB; upon reaching it, metadata recording continues
without new crops. Existing evidence is preserved. No public upload occurs.
- model_class_score and normalized_template_similarity are separate
measures; neither is a calibrated probability of identity.
- Screen containment is recorded as context when available. The current screen
polygon can be a fallback/incorrect candidate, so it does not establish that
the object is physically in the room or on the television.

Operation and enrollment

bash
systemctl --user status sol-object-registry.service
python3 /home/david/random/bin/sol_object_registry.py status

Enroll another named object using a confirmed upright reference image and
its pixel crop (x, y, width, height):

bash
python3 /home/david/random/bin/sol_object_registry.py enroll \
  --camera record-player --label 'chosen name' \
  --image /path/to/upright-reference.png --bbox 100 100 40 50

--camera c615 is also supported. Reference matching is camera-scoped. New
named entries and the initial spider import are persisted in SQLite; a
generated ~/.local/state/sol-vision-camera/named-objects.json projection keeps
the existing renderer and worker in agreement. Do not independently edit that
projection as the registry authority. Raw/secondary overlays are unchanged;
secondary named observations appear in the registry.

Stopping the registry service leaves camera ownership and playback alone.
It refuses to start if /mnt/sol-data is not mounted. The loopback worker API
is port 18919; the existing dashboard proxies only allowlisted read routes.
The viewer is read-only; enrollment is a local CLI operation.

Installed inventory and online comparison

The search covered user services, container inventory, Flatpak applications,
system packages and command-line tools, Python packages, installed model files,
and the prior detector benchmark reports. Installed and relevant:

| Utility | Role and decision |
|---|---|
| OpenCV 4.13 system / 5.0 detector venv | Existing capture/tracking, reference matching, image processing. Reused. |
| OpenVINO 2026.3.1 + YOLO11n ONNX | Bounded CPU generic detector. Reused in the worker. |
| Ultralytics 8.4.150 + YOLOE-11s + MobileCLIP | Installed text/visual prompting option. Benchmarked; not enabled for alerts or substituted for user identities. |
| PyTorch 2.14 CPU / torchvision 0.29 | Installed detector environment; no CUDA inference assumed. |
| MediaPipe ONNX/HOG | Existing person-candidate system; untouched. |
| SmolVLM 500M | Existing scene-description path; prose remains hypothesis rather than database identity. |
| Motion 4.3.2 | Installed isolated diagnostic, disabled after noisy earlier trials. Not activated. |
| Tesseract, FFmpeg, GStreamer | Installed OCR/transport tools; not object registries. |
| Frigate | No running container found; earlier download was paused. Not installed during this change. |

Official options reviewed:

- YOLOE: text/visual-prompt detection
and segmentation. It supports custom categories, but the QuickCam still needs
scene-specific validation. Visual-prompt class IDs require application naming.
- Ultralytics tracking: multiple
tracking backends, including ByteTrack and BoT-SORT. Useful for a higher-rate
continuous tracker, but their IDs alone are not durable real-world identities.
- Frigate: full NVR with
review timelines and tracked objects. Broader than the current registry need;
replacing the existing capture/display stack is unnecessary here.
- SAM 2 and
SAM 3: richer
segmentation/tracking options. No additional weights were installed; current
face-only labeling does not require mask inference.
- Grounding DINO: open-set
text-guided detection, a future alternative requiring local evaluation.
- Norfair: configurable
online detection association. A tracking library, not an object database.

The registry is detector-independent at its ingestion boundary; additional
detectors can submit source-qualified observations without replacing the
catalogue or conflating their scores.

Validation and limits

Initial replay: both installed generic models reported no >=0.35 detections on
the current QuickCam reference. The first C615 download was a warm-up placeholder
and is excluded as a valid camera benchmark. Subsequent fresh C615 live
observations included a bottle candidate; a saved crop shows the bottle. The
QuickCam also produced intermittent chair candidates. These are logged as
provisional, not treated as proof of reliable general recognition.

The low-resolution camera, changes of viewpoint, low light, occlusion, and
reference aging still limit detection. Enrollment helps a known object; it does
not give the classifier universal recognition. Face templates are not human
facial recognition.

41 unit checks passed across the registry and existing named-object, vision,
music, tracking and dashboard suites. Lifecycle checks cover durable IDs,
duplicate samples, stale/offline sources, detector failure, multiple same-class
objects, camera separation, import idempotency, and reacquisition. Live checks
cover both sources, private evidence, viewer rendering, public rejection, and
restart persistence. Raw research artifacts are private under
/mnt/sol-data/sol-stack/object-registry/research/.