75.93%
Top-1 accuracyICW full test set
80.45%
Retrieval mAPAcross unseen identities
94.94%
Average AUC13 PetFace families
60ms
End-to-endTensorRT FP16
Why MeowID
Complementary cues across visual scales.
Individual appearance varies with imaging conditions, making a single representation insufficient to capture all identity cues.
MeowID models facial detail and whole-cat appearance with separately parameterized experts. Facial evidence is prioritized when alignment succeeds; global appearance supplies complementary context and a robust fallback route.
Face-priority routing
Open-set enrollment
Route-specific galleries
Method
Face-priority retrieval with robust fallback.
The system preserves fine-grained facial discrimination while retaining coverage for unconstrained images.
Routing behavior
Select face availability
Face evidence is fused with a whole-cat hint before face-gallery retrieval.
Input photo
Unconstrained image
Aligned face crop
Localized evidence
01
Face expert
Fine-grained facial detail
02
Whole-cat expert
Global appearance
Shared hint fusion
Whole-cat correction
Face gallery
Face-route retrieval
Whole-cat gallery
Fallback retrieval
Identity ranking
Route-specific result
Face localization
Nine landmarks normalize translation, in-plane rotation, and scale into a 256 × 256 aligned crop.
Independent experts
Separate DINOv3 ViT-B/16 encoders produce 512-dimensional face and whole-cat embeddings.
Gated context fusion
A sample-dependent gate contributes global context while preserving the face representation.
Gallery retrieval
Similarity search ranks identities; new individuals can be enrolled without model retraining.
Individual Cats in the Wild
A large-scale benchmark for unconstrained cat retrieval.
ICW is an identity-disjoint benchmark collected from six public adoption and rescue platforms. Multiple observations per identity capture practical variation in pose, viewpoint, background, illumination, camera, and occlusion.
View on Hugging Face View on ModelScope
Results
Improved top-ranked retrieval.
MeowID achieves the best Top-1 accuracy and mAP on the full ICW test set while retaining valid retrieval for all queries without detectable faces.
ICW full set
Best
Top-1 retrieval accuracy
+4.07 percentage points in Top-1 accuracy over PetFace.
Face route
Whole-cat context improves the face route.
Face only74.49%
+4.27 pp
Face + hint78.76%
19.8%Improved
77.2%Unchanged
3.0%Degraded
Ground-truth rank improved for 537 queries and degraded for 82.
Full ICW test set
2,846 queries · 500 unseen identities
Retrieval comparison
| Model | Top-1 ↑ | Top-5% ↑ | Top-50% ↑ | mAP ↑ |
|---|---|---|---|---|
| MegaDescriptor | 56.85 | 89.78 | 99.75 | 65.45 |
| NekoNet | 57.24 | 89.88 | 94.45 | 67.48 |
| PetFace | 71.86 | 90.62 | 94.31 | 78.08 |
| MeowID-Base | 75.93 | 90.72 | 96.52 | 80.45 |
Beyond cats
Generalization across
13 animal families.
On PetFace verification, the MeowID embedding formulation reaches 94.94% average AUC and outperforms the family-specific PetFace-ArcFace baseline across all evaluated families.
94.94AVG. AUC
Deployment
End-to-end inference at 60 ms.
Latency includes decoding, preprocessing, face localization, alignment, embedding extraction, and routing.
PyTorch GPUFP3294.23 ms
ONNX Runtime GPUFP3279.88 ms
TensorRT GPUFP16 · 16.66 img/s60.00 ms
Citation
Cite MeowID.
Use the following reference when building on this work.
@article{hu2026meowid,
title = {MeowID: A Dual-Expert Retrieval System
for Individual Cat Identification},
author = {Hu, Zhangchi and Shang, Yi and Yang, Haocheng
and Hu, Qiwei and Li, Yuzheng},
year = {2026}
}
Open research in individual animal identification
Retrieval beyond