Kellection/dinoV3-ToothVLM-Sonata

DINOv3 ToothVLM: Dental Disease Detection & Periodontal Classification (Sonata)

Models fine-tuned on the Sonata dental dataset, built on a DINOv3 (ViT-B/16) backbone.

License notice: these weights are derived from the DINOv3 pretrained backbone (dinov3_vitb16_pretrain_lvd1689m), which is released under the DINOv3 License (non-commercial). By using these models you agree to the original DINOv3 license terms.

Repository Layout

weights/
├── detection/best_map.pth         # Multi-disease detector (1.1 GB)
└── classification/best_val_acc.pth # Periodontal classifier (661 MB)

---

1. Detector — Multi-Disease Detection (weights/detection/best_map.pth)

DINOv3 ViT-B/16 backbone + YOLOv10-style end-to-end detection head (NMS-free), trained on the Sonata All_Diseases COCO split (mixed training of all disease categories).

Classes (4)

caries · calculus · mouth_ulcer · tooth_discoloration

Performance (COCO bbox, validation)

Metric	Value
mAP@0.5:0.95	0.444
mAP@0.5	0.699
mAP@0.75	0.477
mAP small / medium / large	0.360 / 0.454 / 0.419
Best mAP (training)	0.451

Training setup

Hyperparameter	Value
Backbone	DINOv3 ViT-B/16 (lvd1689m pretrained)
Epochs	50
Batch size	8
Base LR / backbone LR	0.001 / 0.0001
Warmup epochs	5
Unfrozen transformer blocks	6
Loss	E2E (YOLOv10)

Usage

import torch

ckpt = torch.load("weights/detection/best_map.pth", map_location="cpu")
# State dict contains the DINOv3 backbone + v10Detect head.
# See train_detector_sonata.py in the source repo for the exact architecture,
# preprocessing (resize to 1024, normalization) and inference (e2e, no NMS) code.

---

2. Classifier — Periodontal Disease Classification (weights/classification/best_val_acc.pth)

DINOv3 ViT-B/16 backbone + CNN classification head (Conv+BN+ReLU → GAP → Dropout → FC), trained on the Sonata Periodontal_Disease subset with a stratified 70/15/15 train/val/test split.

Classes (3)

gingival_diseases · non_periodontal_disease · periodontitis

Performance (test set, n = 353)

Metric	Value
Accuracy	0.771
Macro-F1	0.772

Per-class F1:

Class	Precision	Recall	F1
gingival_diseases	0.788	0.741	0.764
non_periodontal_disease	0.641	0.750	0.691
periodontitis	0.883	0.838	0.860

Training setup

Same backbone fine-tuning strategy as the detector (frozen backbone, last 6 transformer blocks unfrozen, backbone LR 0.0001).

Usage

import torch

ckpt = torch.load("weights/classification/best_val_acc.pth", map_location="cpu")
# See train_classifier_periodontal.py for the model definition,
# image preprocessing (resize 512, ImageNet normalization) and inference.

---

Citation

If you use these models in your research, please cite the DINOv3 paper:

@article{oquab2025dinov3,
  title={DINOv3: Learning Robust Visual Features without Pretraining},
  author={Oquab, Maxime and others},
  journal={arXiv preprint arXiv:2505.18478},
  year={2025}
}