import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import transforms
from PIL import Image
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(repo_id="DavronSherbaev/person-reid-arcface", filename="arcface_dinov2.pth")

# Load checkpoint
checkpoint = torch.load(model_path, map_location='cpu')

# Create model (see full code in repo)
# model = ArcFaceReID(...)
# model.load_state_dict(checkpoint['model_state_dict'])

# Extract embedding
# embedding = model.get_embedding(image_tensor)

Training Details

Base Model: DINOv2-ViT-B/14 (facebook/dinov2-base)
Loss Function: ArcFace (Additive Angular Margin)
Embedding Dimension: 768
ArcFace Margin: 0.5
Training Classes: 15
Validation Accuracy: 90.12%

Dataset: Custom person crops with bbox augmentation (scales: 90%, 100%, 120%)
Train Samples: ~7,000 (augmented)
Val Samples: 334 (clean, no augmentation)
Epochs: 50
Learning Rate: 3e-6 with cosine annealing
Batch Size: 48
Optimizer: AdamW
Hardware: NVIDIA RTX 3080 Ti

Intended Use

Person re-identification in office/workplace environments. Designed to identify known individuals from camera feeds.

Limitations

Trained on specific office environment
Best for similar lighting/camera conditions
Not suitable for face recognition (uses full body)

License

MIT