Camera-traps-wild-life

Hello there! 😀

This project was born from the idea of applying such algorithms to camera traps located on my family’s property in a remote area of Italy, where wildlife is frequently observed. The goal was to develop a simple application for fixed camera trap systems, where a basic motion detection algorithm could automatically identify subjects to be analyzed by a fine-tuned machine learning model (image-classification).

Example of results applied to camera traps located on my family’s property 👀 (not used for training but only for inference purposes):

Badger	Human	Human

Using grad-cam, we can have some model explainability: we can see how the model correctly classified the fox and the region of the image where the model focused to reach this conclusion is highlighted:

N.B. This is a work in progress, built with limited resources. Community support is always welcome! 💪

How to install

Local virtual environment

We suggest to use PyCharm Community for following steps 2-8.

1. Install Python 3.9: Make sure you have Python installed on your system. You can download it from the official Python website (https://www.python.org/) and follow the installation instructions for your operating system;
2. Clone the repository;
3. Create a virtual environment;
4. Activate the virtual environment;
5. Mark camera_traps folder as root directory;
6. Install project dependencies:
   1. pip install -r requirements.txt
7. Run the commands for further project dependencies:
   1. poetry lock --no-update
   2. poetry install
8. Run main project script:
   1. python camera_traps/main.py

Now you're all set! 🎉 Happy coding! 😄✨

Using Docker

1. Install Docker: Visit the official Docker website (https://www.docker.com/) and follow the installation instructions for your operating system;
2. Clone the repository;
3. Build the Docker image: navigate to the project's root directory and run the following command to build the Docker image:
   1. docker build -t project_name .
4. Run the Docker container: Once the image is built, start a container with the following command:
   1. docker run -it project_name

🚀 This will launch the project within the Docker container! 🐳

---

Model

Currently, the model being used is EfficientNetB0 (https://keras.io/api/applications/), which was implemented to undergo fine-tuning using the custom dataset. The choice of this model was driven by its high accuracy and relatively low number of parameters. More recent series of the same model result in a decrease in computational performance.

The training sessions were conducted using an NVIDIA GPU GeForce 940MX.

Weights

Some of the best weights obtained after fine-tuning are available at the Google Drive link.

Dataset

The dataset used for training is available at the Google Drive link (~ 2.5 GB).

The current dataset has been obtained by combining multiple sources of data available online in order to assemble a dataset of images captured by camera traps in both daytime and nighttime settings. The currently available image classes are as follows:

label	setting	count
None_of_the_above	day	3000
None_of_the_above	night	400
badger	day	955
badger	night	1474
badger	unspecified	18
bear	day	985
bear	night	420
bear	unspecified	779
bird	unspecified	2777
boar	day	1287
boar	night	675
boar	unspecified	775
cat	day	1045
cat	night	935
cat	unspecified	4759
chicken	unspecified	680
cow	day	1351
cow	night	103
cow	unspecified	1138
deer	day	3805
deer	night	2286
deer	unspecified	561
dog	day	1360
dog	night	124
dog	unspecified	3291
fox	day	1408
fox	night	1320
fox	unspecified	8
hare	day	20
hare	night	1262
hare	unspecified	5110
horse	unspecified	62
human	unspecified	2980
squirrel	unspecified	2775
vehicle	unspecified	2829
weasel	day	1907
weasel	night	1119

The dataset folder structure is then organized as follows:

root
 └── dataset
      ├── label1
      │   ├─ image1.jpg
      │   ├─ image2.png                              
      │   └─ ...
      └── label2
          └── ...

The filename of each image is defined as follows:

{referenceNameDataset}_{nameLabel}_{timeCondition}_{progressiveIndex}.jpg

N.B. The underscores are only used as separators, otherwise CamelCase notation has been used.

The timeCondition field can be: 'day', 'night', 'unspecified'.

Useful dataset links:

• NTLNP (wildlife image dataset): https://paperswithcode.com/dataset/ntlnp-wildlife-image-dataset

• CCT20 (subset): https://lila.science/datasets/caltech-camera-traps

• Sheffield: https://figshare.shef.ac.uk/articles/dataset/Badger_datasets_for_image_recognition/8182370/1

• ENA24: https://lila.science/datasets/ena24detection

• LilaMissouri: https://lila.science/datasets/missouricameratraps

• WCS: https://lila.science/datasets/wcscameratraps

• PennFudan: https://www.cis.upenn.edu/~jshi/ped_html/

License

MIT

Contacts

Please open an issue or contact pietro.foini1@gmail.com with any questions.