Skip to content

Use cases

First pedagogical use case

The first AnnoTrain prototype targets the Bioinformatics Tools MSc course at the University of Neuchâtel, for a session titled A First Neural Network on 2 November 2026.

The initial experiment is deliberately simple. Students receive camera-trap images and choose one of:

  • Animal
  • No animal
  • Unsure

With roughly 40 students annotating 30–50 images each, the session generates approximately 1,200–2,000 individual annotations. The pedagogical reveal:

Congratulations — you just created our training dataset.

The annotations are then converted into a consensus dataset and used to train a first binary image classifier. A fallback dataset/model is prepared before the class so the session does not depend entirely on live training or network availability.

Research use cases

AnnoTrain is not hard-coded for animals — the backend understands generic scientific annotation tasks. Potential applications include:

  • camera-trap images
  • roots
  • microscopy
  • cells and nuclei
  • plant organs
  • disease symptoms
  • seeds
  • colonies
  • histological images
  • plates
  • time-lapse experiments
  • scientific videos

Root imaging

If root segments have already been detected, a contributor could click a segment and assign it to Plant 1 / Plant 2 / ... / Plant 6 — an object-level multiclass annotation task.

DeepPlate as an early real scientific model

AnnoTrain is designed to support existing trained models, not only models trained inside the platform. The current DeepPlate U-Net segmentation model is a good candidate for validating this architecture:

Project: DeepPlate
Task: semantic_segmentation
Architecture: U-Net
Input: image
Output: segmentation mask

Two candidate features:

Try model — a researcher uploads a compatible image, AnnoTrain sends an inference job to a compute worker, and the predicted mask is displayed as an overlay.

Use model for annotation — DeepPlate produces preliminary masks; humans Accept / Correct / Reject them, and corrected masks become new validated annotations, closing the loop:

DeepPlate v1 → predictions → human corrections → Dataset v2 → fine-tuning → DeepPlate v2

This complements the teaching use case: camera traps demonstrate binary classification, while DeepPlate demonstrates semantic segmentation and scientific model reuse.