miewid-msv2

Model Details

Model Description

MiewID-msv2 is a feature extractor trained for re-identification using contrastive learning on a large, high-quality dataset of 54 wildlife species - terrestrial and aquatic - including fins, flukes, flanks, faces.

• Model Type: Wildlife re-identification feature backbone
• Model Stats:
  • Params (M): 51.11
  • GMACs: 24.38
  • Activations (M): 91.11
  • Image size: 440 x 440

Model Sources [optional]

<!-- Provide the basic links for the model. -->

• Repository: https://github.com/WildMeOrg/wbia-plugin-miew-id
• Backbone: https://huggingface.co/timm/efficientnetv2_rw_m.agc_in1k

Usage

Intended use is re-identification of individuals from different species by matching against a database of grount-truth samples. Model features can also be used for species classification by retrieval.

Embedding Extraction

import numpy as np
from PIL import Image
import torch
import torchvision.transforms as transforms
from transformers import AutoModel

model_tag = f"conservationxlabs/miewid-msv2"
model = AutoModel.from_pretrained(model_tag, trust_remote_code=True)

def generate_random_image(height=440, width=440, channels=3):
    random_image = np.random.randint(0, 256, (height, width, channels), dtype=np.uint8)
    return Image.fromarray(random_image)

random_image = generate_random_image()

preprocess = transforms.Compose([
    transforms.Resize((440, 440)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

input_tensor = preprocess(random_image)
input_batch = input_tensor.unsqueeze(0) 

with torch.no_grad():
    output = model(input_batch)

print(output)
print(output.shape)

More Examples

View more usage examples at https://github.com/WildMeOrg/wbia-plugin-miew-id/tree/main/wbia_miew_id/examples

Training Details

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

The dataset used for these experiments was a combination of data from Wildbook platforms (multiple users), Happywhale Kaggle competitions multi-species dataset and multiple publicly available datasets. A small subset of data from Wildbook platforms is available at https://lila.science/datasets.

Example images

Evaluation Results

The multi-species model shows improvements over models trained on single species. Moreover the model shows strong generalization ability for majority of the species when trained and evaluated in a leave-one-out fashion.

Detailed results