Automatic annotation

Agouti offers several options to automatically annotate your sequences and add observations. The AI functions are available through Agouti’s user interface, at the click of a button.

In Agouti multiple AI models are available to choose from and most aimed at species classification. The base for each model is a simple classifier that checks if images are blank, contains an animal, or human. If an animal is detected, a second classifier attempts to identify the species in each image. If a sequence contains multiple images these image-level detections are then further processed to yield a sequence-level observation. In this last crucial step, we apply some decision rules to improve the quality of the AI observations.

The result of this classification process is saved in Agouti as a regular observation. The only difference is that it was added by the AI instead of an user, and it’s marked as such. This makes checking the AI observations for errors easier. You don’t need to check all AI observations, but can use a filter to make a selection. For example, you may only be interested in carnivores so check all of those observations, but simply accept all AI observations of ungulates without checks. It’s up to you to decide on a strategy.

As with any AI there are some things to keep in mind. AI classification quality varies and we recommend to at least manually check a selection of AI observations in your project. There is great variation between projects and differences such as image quality, quantity, context, vegetation etc all affect AI performance. Furthermore, AI in itself is rapidly advancing which makes it challenging to keep up with the latest & greatest. We hope to find a balance between AI performance, quality and usability for users.

AI functionality in Agouti is in active development and not perfect. Still, we hope that it’s an useful tool and speeds up the annotation process. We recommend that you use the latest version available for a model whenever possible. Any feedback on the performance of a specific model is welcome. Also, if you have any data to share that we can use to improve a certain model that is highly appreciated.

If you have built your own classification model and would like to use or offer that in Agouti that is possible too. Contact us at agouti@wur.nl to discuss.

How to use the AI?

There are two ways to use the AI functionality in Agouti. The first one works at the deployment level and processing all images in the background. This is ideal if you have multiple deployments and don’t need immediate results. The second option is the realtime AI which is available during manual annotation.

To use the deployment-level AI:

  1. In your project settings, scroll down to automatic annotation and select a model from the list. Details about each model can be found below. This list may be expanded over time, and we regularly release new version of models.

  2. Go to annotate. For each row there is now an additional button: annotate by AI. If you click this button, the deployment is sent to the queue for automatic processing. After clicking, please wait untill the text that indicates your position in the queue appears. This can take up to 2-3 munites for large deployments. Deployments cannot be manually annotated while the AI is working.

  3. When the AI is done, you will see the progress bar is partly green and the status has been updated to ‘AI is done’. You can now manually annotate the remainder of the deployment. You can also use the browse > observations page to browse through the AI observations.

To use use the realtime AI:

  1. In your project settings, scroll down to automatic annotation and select a model from the list.

  2. Open a deployment for manual annotation.

  3. With a sequence open, click the button ‘Analyse’ on the right-hand side of the screen. This sends the sequence to the AI. Depending on the amount of images in the sequence the AI will suggest one or more observations. You can accept, remove or edit the observations as needed. It can take up to 30 seconds to produce output. If the loading bar completes and there are no observations suggested, the AI was not able to produce any output with enough confidence.

Because AI processing takes lots of computing power, we use a queueing mechanism. Depending on the size of the queue, your deployment may be processed immediately or after several hours. If it hasn’t completed after 24 hours, you can contact us to check the status. The realtime AI is still experimental and may not always work quickly. We are continously working to improve the speed and quality of the AI.

AI models available in Agouti

While we develop AI models ourselves, we try to include works of others as well. There are many high quality initiatives around, some of which have kindly shared their models with us so we can offer it in Agouti. Have a model of your own that you would like to use or contribute? Contact us at agouti@wur.nl.

Generic human/blank model (MegaDetector) v5a

Developed by: Microsoft (until 28 April 2023)
Model that supports four classes: animal, human, vehicle and blank. Good option when a specific model is not yet available for your locations. MegaDetector documentation

DeepFaune (France) v05 + v06

Developed by: Institut National d’Ecologie et d’Environnement (INEE) of the CNRS
Model that covers 28 species common to France. Highly recommended for other parts of Western Europe as well. DeepFaune documentation

Western Europe v4a

Developed by: Agouti.eu
This model has been trained with data from Western Europe, mostly the Netherlands, Belgium, Germany and Luxembourg. It covers about 75 taxonomic units. Some species were grouped to a higher level in the taxonomy because they cannot be reliably told apart by the model at this point. For example, Martes or Turdidae.

Europe v5

Developed by: Agouti.eu
Model covering 100+ species of mammal and bird that occur in Europe. The model has been trained with data from across all corners of Europe, with a much wider range and quantity of images than the model for Western Europe. It also leverages a new underlying architecture that works better when there are multiple animals in view. Because this model is a bit more broad than the Western Europe model it may or may not perform better on your data. We recommend trying out both, depending on where oyu are working.

French Guiana v1

Developed by: Agouti.eu
Model that covers 11 species that are common in French Guiana.

India v2

Developed by: Agouti.eu
Model that covers 14 species common to India.

Nepal (vertical) v1

Developed by: Agouti.eu
Model that covers 13 species common to the lowlands of Nepal (Terai). This model is designed for use with camera that are placed facing down vertically.

Panama v2

Developed by: Agouti.eu
Model that covers 19 species common to central Panama.

Southern Africa v3

Developed by: Agouti.eu
Model that covers 35 species that occur in Southern Africa.