Skip to content

Tools for detecting wildlife in aerial images using active learning

License

Notifications You must be signed in to change notification settings

NINAnor/aerial_wildlife_detection

 
 

Repository files navigation

AIDE: Annotation Interface for Data-driven Ecology

AIDE is two things in one: a tool for manually annotating images and a tool for training and running machine (deep) learning models. Those two things are coupled in an active learning loop: the human annotates a few images, the system trains a model, that model is used to make predictions and to select more images for the human to annotate, etc.

More generally, AIDE is a modular Web framework for labeling image datasets with AI assistance. AIDE is configurable for a variety of tasks, but it is particularly intended for ecological applications, such as the acceleration wildlife surveys that use aerial images.

AIDE is primarily developed by Benjamin Kellenberger, supported by the Microsoft AI for Earth program.

Contents

Highlights

  • Powerful: AIDE explicitly integrates humans and AI models in an annotation loop.
  • Fast: AIDE has been designed with speed in mind, both in terms of computations and workflow.
  • Flexible: The framework allows full customizability, from hyperparameters to models to annotation types to libraries. It provides:
    • Support for image classification, point annotations, and bounding boxes (object detection)
    • Many deep learning-based AI models and Active Learning criteria built-in
    • Interfaces for custom AI models and criteria, using any framework or library you want (see how to write your own model).
  • Fully featured: Beyond image labeling and model training, AIDE has management and graphical user/machine performance evaluation tools built-in, right in the web browser, allowing for advanced, manual label quality checks.
  • Modular: AIDE is separated into individual modules, each of which can be run on separate machines for scalability. It even supports on-the-fly addition of computational workers for computationally intensive model training!

AIDE highlights

News

AIDE v2.1 is out

AIDE v2.1 is out! This includes a new interactive installer for Debian/Ubuntu systems as well as a plethora of bug fixes.

Older news

Demo

A demo of AIDE v2 can be accessed here.

This demo allows exploring the annotation front-end with a number of example datasets, including:

Installation and launching AIDE

See here.

AI models in AIDE

Built-in AI models

AIDE ships with a set of built-in models that can be configured and customized:

Label type AI model Model variants / backbones More info
Image labels AlexNet AlexNet paper
DenseNet DenseNet-161 paper
MNASNet MNASNet paper
MobileNet MobileNet V2 paper
ResNet ResNet-18; ResNet-34; ResNet-50; ResNet-101; ResNet-152 paper
ResNeXt ResNeXt-50; ResNeXt-101 paper
ShuffleNet ShuffleNet V2 paper
SqueezeNet SqueezeNet paper
VGG VGG-16 paper
Wide ResNet Wide ResNet-50; Wide ResNet-101 info
Bounding boxes Faster R-CNN with ResNet-50 (PASCAL VOC); with ResNet-50 (MS-COCO); with ResNeXt-101 FPN (MS-COCO) paper, implementation details
RetinaNet with ResNet-50 FPN (MS-COCO); with ResNet-101 FPN (MS-COCO) paper, implementation details
TridentNet with ResNet-50; ResNet-101 (MS-COCO) paper, implementation details
Segmentation masks DeepLabV3+ with modified ResNet-101 (Cityscapes) paper, implementation details

All models can be configured in various ways through the AI model settings page in the Web browser. They all are pre-trained on ImageNet unless specified otherwise. To use one of the built-in models, simply import the requested one to your project through the Model Marketplace in the Web browser and start training/predicting!

Writing your own AI model

AIDE is fully modular and supports custom AI models, as long as they provide a Python interface and can handle at least one of the different annotation and prediction types appropriately. We greatly welcome contributions and are happy to help in the implementation of your custom models!

See here for instructions on implementing custom models into AIDE.

Publications and References

Please cite the following paper if you use AIDE in your work:

Kellenberger, Benjamin, Devis Tuia, and Dan Morris. "AIDE: Accelerating image‐based ecological surveys with interactive machine learning." Methods in Ecology and Evolution 11(12), 1716-1727. DOI: 10.1111/2041-210X.13489.

@article{kellenberger2020aide,
  title={AIDE: Accelerating image-based ecological surveys with interactive machine learning},
  author={Kellenberger, Benjamin and Tuia, Devis and Morris, Dan},
  journal={Methods in Ecology and Evolution},
  volume={11},
  number={12},
  pages={1716--1727},
  year={2020},
  publisher={Wiley Online Library}
}

If you use AIDE, we would be happy to hear from you! Please send us an E-mail with a little bit of info about your use case; besides getting to know the fellow usership of our software, this also enables us to provide somewhat more tailored support for you if needed. Thank you very much.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

About

Tools for detecting wildlife in aerial images using active learning

Resources

License

Code of conduct

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 50.6%
  • HTML 22.4%
  • JavaScript 21.8%
  • Shell 2.8%
  • CSS 1.9%
  • Dockerfile 0.3%
  • PLpgSQL 0.2%