Image Classification Data Annotation

In this section we will be annotating robot image datasets (e.g., teleop image sequences) for ML training. We will be fine tuning a YOLO (You Only Look Once) vision model with our purple and green ball image datasets.

Recording Images

Follow the instructions under Camera Operations - Recording Image Files for launching the recorder.

When you record images, the recorder will automatically create a timestamped folder with those images in /home/rcr/teleop_data. It creates a new folder each time you start the recorder.

Image Directory Structure

To help manage your image annotation projects, a good practice is to create a separate directory and move the image files you want to label into it.

Make an central images folder on the Pi

On the Pi Raspberry Pi

mkdir -p "/home/rcr/teleop_data/images"

Move images you want to label into that folder

On the Pi Raspberry Pi

# Replace session_YYYY-MM-DD_HH-MM-SS with your actual session folder name
mv /home/rcr/teleop_data/session_*/*.jpg /home/rcr/teleop_data/images/

Label Studio

We have preloaded a software tool called Label Studio onto the Raspberry Pi for data annotation. You can use it by launching a webserver on the Pi and then use a web browser on your development machine to access the Label Studio UI. There you can create bounding-box/classification labels (e.g., purple whiffle, green whiffle), using the image files directly on your Pi with local storage.

Alternatively, there are many cloud annotation software suites available which are continuing to improve and get more powerful. We have also included instructions for Roboflow if you do not like working with Label Studio.

🔗 Full Label Studio Guide

Consolidate image files

Make sure you have followed instructions above for moving your image recording files into a central location. To verify, running ls ~/teleop_data/images on the Pi should not return an error.

Make annotation folder on the Pi

On the Pi Raspberry Pi

mkdir -p "/home/rcr/teleop_data/annotations"

Set required environment variables for your terminal session

On the Pi Raspberry Pi

export LABEL_STUDIO_LOCAL_FILES_SERVING_ENABLED=true
export LABEL_STUDIO_LOCAL_FILES_DOCUMENT_ROOT="/home/rcr/teleop_data"

Verify they are set with:

env | grep LABEL_STUDIO

(Optional) Make env vars persist

This is helpful so you don’t have to manually set the environment variables every time. Once you have done this you can skip this step in the future.

On the Pi Raspberry Pi

nano /home/rcr/.profile

Add to the end:

export LABEL_STUDIO_LOCAL_FILES_SERVING_ENABLED=true
export LABEL_STUDIO_LOCAL_FILES_DOCUMENT_ROOT="/home/rcr/teleop_data"

Close/reopen terminal (or source /home/rcr/.profile) and restart Label Studio.

Verify they are set with:

env | grep LABEL_STUDIO

Start Label Studio

On the Pi Raspberry Pi

cd /home/rcr/LabelMaker
source bin/activate
label-studio start -p 8080 -b

Open the UI

On your development computer 🖥️

  • In a browser on your laptop: http://192.168.1.n:8080 (replace n with your robot’s number / the Pi’s IP last octet)

  • Create a Label Studio account

  • Create a New Project (template: Object Detection with Bounding Boxes)

  • In the template setup for your annotation template, delete the existing class names and add your own.

  • In order to merge all the annotated datasets from the whole class, our conventions must align. Add Purple Ball as the first class and then Green Ball as the second class.

  • You can use the color picker to choose an appropriate color for each class. Then click Save.

Configure Storage in the Project

On your development computer 🖥️

Add Source storage (images to label): Project → Settings → Cloud Storage → Add Source Storage → Local files

  • Storage Title Anything you want here (required)

  • Absolute Local path: /home/rcr/teleop_data/images

  • (Optional) File mask/regex: .*\.(jpg|jpeg|png)$

  • Test Connection

  • Next

  • Import method: Files ← important for images (not “Tasks”)

  • Click Next → Save → Sync

If you see UnsupportedFileFormatError ... Only .json/.jsonl/.parquet... you picked Tasks. Edit storage and switch Import method to Files.

Add Target storage (where annotations are saved): Project → Settings → Cloud Storage → Add Target Storage → Local files

  • Storage Title Anything you want here (required)

  • Absolute Local path: /home/rcr/teleop_data/annotations

  • Path leave blank

  • Save

Label your images

On your development computer 🖥️

Export / Share your annotations

On your development computer 🖥️

  • If Sync on submit is on, results are already in /home/rcr/teleop_data/annotations

  • Otherwise: go to Target Storage and click Sync

  • Zip them for handoff:

    On the Pi Raspberry Pi

    cd "/home/rcr/teleop_data"
zip -r annotations.zip annotations/

Submit your files to the group dataset

  • If you are in the RCR lab - let us know and we will grab the annotated zip file from your ip address.

Quick troubleshooting

  • Broken thumbnails: open DevTools → Network on a task; image URLs should look like /data/local-files/?d=images/<file>

    • 404 → check LABEL_STUDIO_LOCAL_FILES_DOCUMENT_ROOT and path mapping

    • 403 → permissions (ensure read/execute on folders; write on /home/rcr/teleop_data/annotations)

  • Cursor opens localhost automatically: that’s just port-forwarding, its possible that port forwarding might cause an issue so you could directly visit this your Pi IP address (http://192.168.1.n:8080)

  • Verify env vars before start:

    On the Pi Raspberry Pi

    env | grep LABEL_STUDIO

    • Try multiple browsers Since you are accessing the Label Studio UI via a browser, its possible that cross operating system / cross browser issues may occur. If you are experiencing issues with Label Studio you could try switching browsers.

Additional Troubleshooting

  • Broken images not re-syncing

    On your development computer 🖥️

    If you fixed your configuration but tasks/images still won’t refresh, try deleting all tasks and re-syncing the storage:

    1. Go to Projects → Your Project.

    2. In the Data Manager, Select All tasks.

    3. In the Actions dropdown, choose Delete Tasks.

    4. Return to Settings → Cloud Storage → Source (Local files) and click Sync again.

  • Nuclear option (reset Label Studio state)

    On the Pi Raspberry Pi

    Label Studio stores app state (including your login) in a local SQLite database under /home/rcr/.local/share/label-studio/. You can wipe and start fresh if something is corrupted. This does not delete your image/annotation folders under /home/rcr/teleop_data/....

    Stop Label Studio

    pkill -f label-studio

    Remove LS state (DB + caches) and recreate the folder

    rm -rf /home/rcr/.local/share/label-studio
mkdir -p /home/rcr/.local/share/label-studio

    Restart Label Studio with your env vars set (see setup above)

    cd ~/LabelMaker && source bin/activate
export LABEL_STUDIO_LOCAL_FILES_SERVING_ENABLED=true
export LABEL_STUDIO_LOCAL_FILES_DOCUMENT_ROOT="/home/rcr/teleop_data"
label-studio start -p 8080 -b

Roboflow

If Label Studio on the Pi is uncooperative and continues to block you, you can annotate in the cloud with Roboflow. First, copy your images off the Pi to your computer, then upload to Roboflow.

Roboflow is a cloud based platform with a free tier which include 1,000 public annotations. Robflow also has other powerful integrated ML tools to help automated your data labeling process with predictive labeling which go beyond the capabilities of Label Studio.

Copy images from the Pi to your computer

Consolidate image files Make sure you have followed instructions above for moving your image recording files into a central location.

On your development computer 🖥️

macOS / Linux / Windows WSL (Terminal):

# Replace `n` with your robot/Pi's last IP octet and adjust the destination path if desired
scp -r rcr@192.168.1.n:/home/rcr/teleop_data/images "/Users/$USER/Downloads/teleop_images"

Windows (PowerShell with OpenSSH):

# Replace n with your Pi's last IP octet and YourName with your Windows username
scp -r rcr@192.168.1.n:/home/rcr/teleop_data/images "C:\Users\YourName\Downloads\teleop_images"

Prefer a GUI? Use WinSCP: create a new SFTP connection to rcr@192.168.1.n, navigate to /home/rcr/teleop_data/images, and drag the folder to your PC.

Upload & Annotate in Roboflow

On your development computer 🖥️

  • Roboflow Annotate Docs

  • Create an account at Roboflow.com

  • Create a New Project (e.g., Object Detection for bounding boxes).

  • Upload your copied images folder.

  • Use the browser-based annotator to draw boxes and assign labels (e.g., purple ball, green ball).

  • When done, you can Export in your preferred training format (e.g., YOLO) from Roboflow’s dataset settings.