BVRA / MegaDescriptor-L-224

Model card for MegaDescriptor-L-224

A Swin-L image feature model. Supervisely pre-trained on animal re-identification datasets.

Model Details

• Model Type: Animal re-identification / feature backbone
• Model Stats:
  • Params (M): 228.6
  • Image size: 224 x 224
  • Architecture: swin_large_patch4_window7_224
• Paper: WildlifeDatasets_An_Open-Source_Toolkit_for_Animal_Re-Identification
• Related Papers:
  • Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
  • DINOv2: Learning Robust Visual Features without Supervision
• Pretrain Dataset: All available re-identification datasets --> https://github.com/WildlifeDatasets/wildlife-datasets

Model Usage

Image Embeddings

import timm
import torch
import torchvision.transforms as T

from PIL import Image
from urllib.request import urlopen

model = timm.create_model("hf-hub:BVRA/MegaDescriptor-L-224", pretrained=True)
model = model.eval()

train_transforms = T.Compose([T.Resize(224), 
                              T.ToTensor(), 
                              T.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])]) 

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

output = model(train_transforms(img).unsqueeze(0))  # output is (batch_size, num_features) shaped tensor
# output is a (1, num_features) shaped tensor

Citation

@inproceedings{vcermak2024wildlifedatasets,
  title={WildlifeDatasets: An open-source toolkit for animal re-identification},
  author={{\v{C}}erm{\'a}k, Vojt{\v{e}}ch and Picek, Lukas and Adam, Luk{\'a}{\v{s}} and Papafitsoros, Kostas},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={5953--5963},
  year={2024}
}