Why AnnoTrain

The problem

Scientific image analysis often involves several disconnected tools: one application to annotate images, another script to prepare datasets, another framework to train a neural network, notebooks to evaluate it, and yet another platform to distribute the resulting model.

For students and non-specialists, this fragmentation also makes deep learning appear as a black box.

AnnoTrain aims to connect these steps in a single, understandable workflow:

Human annotation → Dataset → Training → Evaluation → Model → Sharing → Reuse → Improvement

A researcher should eventually be able to:

1. create a project;
2. upload scientific images;
3. define what should be annotated;
4. share a simple annotation link;
5. collect annotations from students, colleagues, or citizen scientists;
6. build a quality-controlled dataset;
7. optionally train a baseline deep-learning model;
8. inspect its predictions and errors;
9. ask humans to correct uncertain predictions;
10. create an improved dataset/model version;
11. publish or download the dataset and model.

A contributor should have an even simpler experience:

Open link → annotate → next image.

An account should not be required for simple public annotation projects, at least initially.

Education is a core feature

Education is not an optional addition to AnnoTrain — it is one of its four pillars, alongside participatory annotation, deep learning, and open science.

AnnoTrain should help make deep learning less of a black box by exposing the complete process:

images → human labels → ground truth → dataset split → training →
predictions → evaluation → errors → improvement

Students should be able to see that a neural network does not magically know the correct answer: its behavior depends on the data humans provide.

The interface should eventually make concepts such as the following tangible:

• labels and classes
• training, validation, and test sets
• class imbalance
• annotation errors and human disagreement
• consensus
• probabilities, loss, epochs, overfitting
• accuracy, precision, recall, F1 score
• confusion matrices
• segmentation metrics such as IoU and Dice
• transfer learning
• model uncertainty
• human-in-the-loop learning

A particularly useful comparison is human consensus vs. model confidence, for example:

Source	Estimate
Humans	54% Animal
Model	P(Animal) = 0.51

Cases like this can be used to discuss the meaning and limitations of "ground truth".

Positioning relative to existing tools

AnnoTrain does not attempt to reproduce every feature of established annotation or ML platforms. Relevant existing tools include:

• Zooniverse — citizen-science workflows
• CVAT — advanced computer-vision annotation
• Label Studio — configurable annotation workflows
• Roboflow — integrated computer-vision data/training workflows
• FiftyOne — dataset/model visualization and exploration

AnnoTrain's distinctive objective is the combination of citizen science + scientific annotation + deep learning + education + open science. The goal is not to outperform specialized tools in every individual category — for complex annotation editors such as segmentation, existing components or libraries should be reused where practical rather than building a complete image editor from scratch.

MVP philosophy

AnnoTrain is built incrementally. The first milestone answers one question:

Can approximately 40 students simultaneously open a URL, annotate camera-trap images, and create a usable dataset?

The first functional workflow is therefore intentionally minimal:

Open URL → Show image → Animal / No animal / Unsure → Save annotation → Show next image

No segmentation editor. No tracking. No automated publication. No complicated authentication. No Kubernetes. No large AutoML system — those come later, once the core loop works. See the Roadmap for the full progression.