README.md

---
title: Age Estimation Demo
emoji: 🚀
colorFrom: blue
colorTo: purple
sdk: gradio
python_version: 3.12
app_file: app.py
license: apache-2.0
tags:
  - age-estimation
  - gender-classification
  - face-analysis
  - vision-transformer
  - dinov3
  - coral-ordinal-regression
pipeline_tag: image-classification
---

# FaceAge-DINOv3

Age and gender estimation from face crops using **DINOv3-ViT-L** backbone with CORAL ordinal regression.

## Performance (LAGENDA benchmark)

| Model | MAE ↓ | CS@5 ↑ | Gender Acc ↑ |
|-------|--------|--------|-------------|
| MiVOLO v2 [face+body] (paper) | 3.650 | 74.48% | 97.99% |
| MiVOLO v2 [face+body] (measured on the public model) | 3.859 | 76.5% | — |
| MiVOLO v2 [face-only] (measured on the public model) | 3.941 | 75.6% | — |
| **FaceAge-DINOv3 (face-only)** | **3.760** | — | — |

Trained on:Our collection data.

## Architecture

```
Face [B, 3, 224, 224]
    ↓
DINOv3-ViT-L/16  (307M params, pretrained on LVD-1.68B)
    ↓ pooler_output
[B, 1024]
    ↓ LayerNorm → Linear(1024→512) → GELU → Dropout
[B, 512]
    ├── age_head:    Linear(512, 100) → CORAL → age ∈ [0, 100]
    └── gender_head: Linear(512, 2)  → softmax → {female, male}
```

**CORAL ordinal regression**: age = Σ σ(logit_k) for k=0..99. Exploits the ordinal structure of ages (25 < 26 < 27) for better calibration than standard cross-entropy.

## Usage

```python
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

processor = AutoImageProcessor.from_pretrained("trungthanhtran/faceage-dino")
model = AutoModel.from_pretrained("trungthanhtran/faceage-dino",
                                   trust_remote_code=True)
model.eval()

# Input: 224×224 face crop (already cropped, no detection needed)
image = Image.open("face_crop.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")

import torch
with torch.no_grad():
    outputs = model(**inputs)

age    = outputs.age_output.item()
gender = "male" if outputs.gender_class_idx.item() == 1 else "female"
conf   = outputs.gender_probs[0, outputs.gender_class_idx.item()].item()

print(f"Age    : {age:.1f}")
print(f"Gender : {gender}  (conf={conf:.2f})")
```

## ONNX (no PyTorch needed)

The model is also available as a single-file ONNX for CPU deployment:

```bash
pip install onnxruntime numpy pillow
python infer_onnx.py --onnx faceage_dino_fp32.onnx --image face.jpg
```

ONNX is ~3-4× faster on CPU than the PyTorch model and requires no GPU.

## Benchmark against MiVOLO v2

```bash
python infer_onnx.py \
    --onnx   faceage_dino_fp32.onnx \
    --lagenda_dir   data/lagenda \
    --annotation_csv lagenda_test.csv \
    --batch_size 256
```

## Training

Multi-phase fine-tuning on DINOv3-ViT-L:

| Phase | Backbone | LR | Data |
|-------|----------|-----|------|
| 1 | Frozen (all 24 blocks) | 1e-3 | Our collection  786k faces |
| 2 | Top 4 blocks unfrozen | 1e-4 | Same |
| 3 | All blocks unfrozen | 3e-5 | Same |
| 4 | All blocks | 3e-6 | Our collection  4M faces, age reweighting |

Age group reweighting (Phase 4): 36-50 ×2.0, 51-65 ×1.5, 66-100 ×3.0 to improve accuracy on older faces.

## Citation

If you use this model, please cite:

```bibtex
@misc{faceage-dino-2026,
  title  = {FaceAge-DINOv3: Age and Gender Estimation with DINOv3-ViT-L},
  author = {Trung Thanh Tran},
  year   = {2026},
  url    = {https://huggingface.co/trungthanhtran/faceage-dino}
}
```

Also cite the backbones and datasets used:
- DINOv3: Meta AI, "DINOv3: Scaling Up Vision Foundation Models", 2025
- LAGENDA: Bhuiyan et al., 2023
- MiVOLO: Kuprashevich & Tolstykh, arXiv 2307.04616