Multimodal Annotation Demo – Image Project Overview

This project shows part of the work a remote Data Annotator would do for an AI research team: setting up an image‑classification project, labeling images with clear rules, and exporting annotations for model training or evaluation. Label Studio is used as the main annotation tool for this demo.​

Project structure

• data/images/ – sample images taken by the annotator.

• labels/image_annotations.csv – exported labels from Label Studio in CSV format.​

• guidelines/labeling-guidelines.md – written labeling instructions used while annotating.​

• screenshots/ – step‑by‑step screenshots showing the annotation process.

• env/ – optional Python virtual environment (not committed to Git).

Annotation task

• Tool: Label Studio, running locally in a Python virtual environment.​

• Project: Image Annotation – Demo.

• Goal: Classify everyday photos into safety and content categories used by content‑moderation and perception models.

• Labels used: Adult content, Weapons, Violence, Vehicle, Airplane, Airport Terminal.

• Mode: Single‑label classification – one primary label per image, even if multiple concepts appear.​

Labeling workflow

1. Start Label Studio with label-studio start from the project venv to open the web UI at http://localhost:8080.​

2. Create the Image Annotation – Demo project using the Image Classification template and configure the labels above.​

3. Import photos from data/images/ into the project.

4. For each image, apply the guidelines from guidelines/labeling-guidelines.md and click Submit to save the annotation.​

5. After labeling, return to the task list and use Export → CSV to download annotations and save them as labels/image_annotations.csv.​

Labeling guidelines (summary) Full details live in guidelines/labeling-guidelines.md. In short:

• Choose the single label that best matches the main visible subject or risk in the image.

• Use Vehicle for real full‑size vehicles (cars, trucks, etc.), and Airplane when an aircraft is the clear primary subject.

• Do not label toys or decorations as vehicles; do not infer unseen weapons or violence.

• If an image does not match any label (for example, mugs on a table), submit it with no boxes checked.

Screenshots for walkthrough The screenshots/ folder contains:

• Login and project‑dashboard views for Label Studio.

• Label‑setup configuration screen with the custom label set.

• Example labeling screens for several images.

• Task list showing completed annotations and the export button.

• These images help other readers understand the exact UI steps taken without needing to run the tool themselves.​

Evaluation

AI output review: The evaluation/ai_output_review.csv file simulates how a data annotator would review model predictions against human labels and record correctness and feedback.

Author: Erik V (Data Annotator)