makactuary/simba-or-not

---

license: cc-by-sa-4.0

library_name: sklearn

tags:

• image-classification

• cats

• fun

• timm

datasets:

• timm/oxford-iiit-pet

base_model: timm/convnext_tiny.fb_in22k_ft_in1k

---

Is this Simba? 🐾

A one-cat detector for Simba, an orange-and-white long-haired cat: neither Jedi nor Sith, more of a gray-area master of the sofa. A logistic-regression head on frozen ConvNeXt-Tiny features, trained on 42 of the owner's photos of Simba against 2371 cat photos from the Oxford-IIIT Pet dataset. A personal, playful model: it answers one question about one cat.

Evaluation

Simba's photos were grouped by the day they were taken (20 days), so every test fold contains only days the model had not seen (5-fold group CV; near-identical burst shots never span train and test).

• Simba recognised on unseen days: 41 of 42 photos (mean recall 0.98, lowest fold 0.95).

• Other cats wrongly called Simba: 1 of 2371 (0.04%), a Maine Coon (1 of 200 Maine Coons), the closest-looking breed.

Limitations

• Every Simba photo was taken in the same home, while the other cats come from many places: part of what the model learned may be the home's carpet, sofa or light. It has not been tested on Simba photographed elsewhere, or on other orange cats in the same home.

• 42 photos from 20 days is a small sample; treat the numbers as indicative.

Usage

python predict.py photo.jpg → ('Simba', 0.97) (needs timm, torch, scikit-learn 1.9.1, huggingface_hub, Pillow).

Data and licence

Negative examples: Oxford-IIIT Pet (Parkhi et al., 2012, CVPR), CC BY-SA 4.0, so this model is shared under CC BY-SA 4.0. Simba's photos belong to his owner and are not published; only the trained head is. Backbone: timm/convnext_tiny.fb_in22k_ft_in1k (Apache-2.0).

Code: hf_image_models.py in QMASE.