Tools for training and running detectors and classifiers for wildlife images collected from motion-triggered cameras.
|size (curr.)||1374794 kB|
This repo contains the tools for training, running, and evaluating detectors and classifiers for images collected from motion-triggered camera traps. The core functionality provided is:
- Data parsing from frequently-used camera trap metadata formats into a common format
- Training and evaluation of detectors, particularly our “MegaDetector”, which does a pretty good job finding terrestrial animals in a variety of ecosystems
- Training and evaluation of species-level classifiers for specific data sets
- A Web-based demo that runs our models via a REST API that hosts them on a Web endpoint
- Miscellaneous useful tools for manipulating camera trap data
- Research experiments we’re doing around camera trap data (i.e., some directories are highly experimental and you should take them with a grain of salt)
Classifiers and detectors are trained using TensorFlow.
This repo is maintained by folks in the Microsoft AI for Earth program who like looking at pictures of animals. I mean, we want to use machine learning to support conservation too, but we also really like looking at pictures of animals.
How we work with ecologists
We work with ecologists all over the world to help them spend less time annotating images and more time thinking about conservation. You can read a little more about how this works on our AI for Earth camera trap collaborations page.
You can also read about what we do to support camera trap researchers in our recent blog post.
This repo does not directly host camera trap data, but we work with our collaborators to make data and annotations available whenever possible on lila.science.
This repo does not extensively host species classification models, though we will release models when they are at a level of generality that they might be useful to other people. But…
Speaking of models that might be useful to other people, we have trained a one-class animal detector trained on several hundred thousand bounding boxes from a variety of ecosystems. Lots more information – including download links – on the MegaDetector page.
Here’s a “teaser” image of what detector output looks like:
Image credit University of Washington.
For questions about this repo, contact email@example.com.
This repo is organized into the following folders…
Code for hosting our models as an API, either for synchronous operation (e.g. for real-time inference or for our Web-based demo) or as a batch process (for large biodiversity surveys).
Code for training species classifiers on new data sets, generally trained on crops generated via an existing detector. We’ll release some classifiers soon, but more importantly, here’s a tutorial on training your own classifier using our detector and our training pipeline.
Oh, and here’s another “teaser image” of what you get at the end of training a classifier:
- Converting frequently-used metadata formats to COCO Camera Traps format
- Creating, visualizing, and editing COCO Camera Traps .json databases
- Generating tfrecords
Source for the Web-based demo of our MegaDetector model (we’ll release the demo soon!).
Code for training and evaluating detectors.
Ongoing research projects that use this repository in one way or another; as of the time I’m editing this README, there are projects in this folder around active learning and the use of simulated environments for training data augmentation.
Random things that don’t fit in any other directory. Currently contains a single file, a not-super-useful but super-duper-satisfying and mostly-successful attempt to use OCR to pull metadata out of image pixels in a fairly generic way, to handle those pesky cases when image metadata is lost.
We use conda to manage our Python package dependencies. Conda is a package and environment management system. You can install a lightweight distribution of conda (Miniconda) for your OS via installers at https://docs.conda.io/en/latest/miniconda.html.
Utility and visualization scripts
The required Python packages for running utility and visualization scripts in this repo are listed in environment.yml. To set up your environment for these scripts, in your shell, navigate to the root directory of this repo and issue the following command to create a virtual environment via conda called
cameratraps (specified in the environment file) and install the required packages:
conda env create --file environment.yml
For unix users, you need to have gcc installed in order to compile the pip packages. If you do not already have gcc installed, run the following command before creating the conda environment:
sudo apt update sudo apt install build-essential
Machine learning scripts
Scripts that execute machine learning code – specifically, scripts in the folders
classification – require additional depdendencies. In particular, the
detection/run_tf_detector*.py scripts should use environment-detector.yml to set up the environment, as follows:
conda env create --file environment-detector.yml
This environment file allows any TensorFlow version from 1.9 to 1.15 to be installed, but you may need to adjust that version for your environment. Specifically, if you are running on an Azure Data Science Virtual Machine (which has CUDA 10.1 as of the time I’m writing this), you may receive a CUDA error, in which case you should change the line:
- tensorflow-gpu>=1.9.0, <1.15.0
…before creating your environment.
If you run into an error while creating either of the above environments, try updating conda to version 4.5.11 or above. Check the version of conda using
To enter the conda virtual environment at your current shell, run:
conda activate cameratraps
…or, if you used the environment-detector.yml file above:
conda activate cameratraps-detector
You should see
(cameratraps) prepended to the command line prompt. Invoking
jupyter notebook will now be using the interpreter and packages available in this virtual env.
To exit the virtual env, issue
Add additional packages
If you need to use additional packages, add them to the environment file and run
conda env update --name cameratraps --file environment.yml --prune
conda env update --name cameratraps-detector --file environment-detector.yml --prune
In some scripts, we also assume that you have the AI for Earth utilities repo (
ai4eutils) cloned and its path appended to
PYTHONPATH. You can append a path to
PYTHONPATH for the current shell session by executing the following on Windows:
You can do this with the following on Linux:
Adding this line to your
~/.bashrc (on Linux) modifies
We also do our best to follow Google’s Python Style Guide, and we have adopted their
pylintrc file, with the following differences:
- indent code blocks with 4 spaces (instead of 2)
Gratuitous pretty camera trap picture
Image credit USDA, from the NACTI data set.
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 cla.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., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This repository is licensed with the MIT license.