Pet Recognition

Pet Recognition tells your individual pets apart. With it enabled, two dogs in your library become Rocky and Luna — two people you can name, rename, merge and browse — instead of a single shared "dog" entry.

It builds on Pet Detection: detection finds the animal in the photo, recognition decides which animal it is.

Dogs and cats only

Only dogs and cats are recognized as individuals — and they are now the only two species Pet Detection records at all. Birds, horses, sheep and cows used to land in a shared per-species entry that the People page never listed; those were write-only rows, so detection stopped writing them.

This is a limit of the model, not an oversight: it is trained on dog and cat identities, so it has no basis for telling one bird apart from another — which is exactly why recording those species had nowhere useful to go.

Enabling it

Pet Recognition is off by default, and turning it on is not enough on its own — see Reprocessing your library below.

1. Go to Administration > Settings > Machine Learning Settings > Pet Recognition.
2. Enable Enable pet recognition.
3. Save.

Pet Detection must also be enabled, otherwise no new photos are scanned for animals at all and recognition has nothing to work with. The settings page warns you when detection is off.

From that point on, newly uploaded photos go through the individual pipeline. Photos already in your library keep whatever they had — so until you reprocess, the library is in a mixed state: old photos still attached to the shared "dog" entry, which is not listed in the People section, while new ones appear there as named-able individuals.

Reprocessing your library

This deletes data

Reprocessing deletes every pet person and every pet embedding you have, including any names you gave them and the copies that were projected into shared spaces. There is no undo. The names are gone even though the photos are not.

To apply recognition to photos that were uploaded before you enabled it:

1. Go to Administration > Jobs.
2. On the Pet Recognition queue, press Reset.
3. Confirm the dialog.

Gallery then deletes all pet people and embeddings and re-runs pet detection over the whole library, embedding and clustering as it goes. On a large library this takes a while — it re-runs the detector on every asset.

Reset while pet detection is disabled

The deletion always happens, but the rebuild is gated on Pet Detection being enabled — that is what the confirmation dialog means by "reprocessing only runs while pet detection is enabled".

So if you reset while detection is off, your library ends up with no pets at all until you re-enable detection and reset again. Enable Pet Detection first.

Naming your pets

Recognized pets appear on the People page with a paw badge, exactly like recognized faces. Everything you already do with people works:

• Click an unnamed pet and give it a name.
• Merge two entries when the same pet was clustered twice.
• Hide a pet you don't want on the People page.
• Open a pet to browse every photo it appears in.

Pets propagate into shared spaces the same way people do, so a space's People page shows the pets in that space's photos.

Model options

Three models are available, all producing the same 512-dimension embedding, so you can switch between them without a schema change. Larger models are more accurate and slower, and the download is not trivial.

Model	Backbone	Download	Top-1 (dogs)	Top-1 (cats)
pet-recognition-small	DINOv2-S, 22M	~89 MB	0.535	0.913
pet-recognition-base	DINOv2-B, 86M	~348 MB	0.612	0.916
pet-recognition-large	DINOv2-L, 300M	~1.2 GB	0.672	0.915

The default is base. Accuracy is measured on held-out identities — pets the model never saw during training — from the Dogs-World and Cat Individual Images test sets.

Cats score about the same on all three, so if your library is mostly cats there is little reason to pay for a bigger model. Dogs are where the larger models earn their size.

Switching the model deletes your named pets

Every model produces its own embedding space, so embeddings from one model are meaningless to another. Changing the model therefore deletes all pet people and their embeddings and reprocesses your entire library, exactly like a reset. The admin page asks you to confirm before saving.

Species entries left over from before detection narrowed to cats and dogs are pure detector output and are not model-coupled, so they survive a model switch untouched. Nothing creates new ones.

Configuration

Under Administration > Settings > Machine Learning Settings > Pet Recognition:

Maximum recognition distance

How far apart two pets can be in embedding space and still be treated as the same individual, from 0.1 to 2. Default 0.55.

• Lower it if Gallery is merging two different pets into one — common with two dogs of the same breed and colour.
• Raise it if the same pet keeps showing up as several separate individuals.

Changing this affects future recognition jobs; existing assignments stay as they are until you reset.

Minimum recognized faces

How many similar pet detections must exist before Gallery creates a named-able individual, from 1 to 1000. Default 1.

Increasing it makes recognition more precise, at the cost of pets with only a photo or two never being grouped into an individual at all. The default of 1 suits most home libraries, where a pet may only appear in a handful of photos.

Pet Recognition model

See Model options above. Changing this deletes and rebuilds everything.

How it works

1. Detection — the pet detector finds animals and their bounding boxes (see Pet Detection).
2. Embedding — for dogs and cats, the same machine learning request crops the animal, resizes it to 224x224, and runs it through a frozen DINOv2 backbone plus a trained linear projection, producing a 512-dimension L2-normalized embedding. These are stored in a dedicated pet_search table, entirely separate from human face embeddings.
3. Clustering — a PetRecognition job runs a nearest-neighbour search over that table, scoped to the owner. A match within the recognition distance joins the existing individual; otherwise a new one is created.
4. Shared spaces — the pet is projected into every space its photo belongs to, the same path human faces already take.

Human and pet embeddings never mix: they live in different tables, are searched by different queries and are queued on different job queues, so a pet can never be clustered onto a person or vice versa.

Recognition also runs nightly, picking up any pet detections that have not yet been assigned to an individual.