theavenger/cat-reid-dinov3

Cat ReID — DINOv3 Fine-tuned

Fine-tuned DINOv3 ViT-B/16 for cat re-identification.

Usage

import torch
from ai.feature_extractor import VisionFeatureExtractor

extractor = VisionFeatureExtractor(device="cpu", checkpoint="finetuned_best.pt")
feature = extractor.extract(image)  # 256-dim normalized embedding

Model

• Backbone: DINOv3 ViT-B/16 (768-dim)
• Head: Linear(768 → 256) + L2-norm
• Training: Triplet margin loss with batch-hard mining
• Classes: 131 cats