Enhancing Active Learning: Uploading Data to Roboflow from Raspberry Pi using the API (Part 7)

Building Your Own Real-Time Object Detection App: Roboflow(YOLOv8) and Streamlit in Raspberry

Introduction

Active learning aims to select the most informative samples from a large pool of unlabeled data to be labeled by usually a human annotator to improve the model’s performance.

In the context of Raspberry Pi, active learning in this project will be collecting data and uploading this new data in Roboflow to increase model performance. It can be particularly useful when working with small datasets, which is common in embedded systems like the Raspberry Pi, where storage and computational resources may be limited.

To Achieve this I’ll use Roboflow API to upload images to our existing dataset on the Roboflow platform. You can upload images to Roboflow projects using the web interface, Python SDK, REST API, and CLI.

First we have to have set up our Raspberry Pi check my tutorial post how to set up a Raspberry Pi. To collect data I’m using Buster version of Raspberry Pi OS, then we have to set up the camera, if you are using an official camera from Raspberry Pi check my tutorial how to set up the camera with the OS version that I mentioned.

Once we have everything ready the first thing we need is create a Python code to automate as we did in Windows how to take photos, the code will create a folder with the name of the labels list we want, and for each folder will captures images with the camera the times we asked and as we did in Windows will appear a windows showing us what the camera see, let’s see the code:

import time
import picamera
from PIL import Image
import os

def get_image_size(image_path):
    with Image.open(image_path) as img:
        return img.size

def take_pictures(label, num_pictures, interval_seconds, output_folder):
    folder_path = os.path.join(output_folder, label)
    if not os.path.exists(folder_path):
        os.makedirs(folder_path)

    with picamera.PiCamera() as camera:
        # Start the preview
        camera.start_preview()

        for i in range(1, num_pictures + 1):
            # Generate the file name for the picture
            file_name = os.path.join(folder_path, f"{label}_picture{i}.jpg")

            # Capture the picture
            camera.capture(file_name)

            # Get the size of the captured image
            image_size = get_image_size(file_name)
            print(f"{label} - Picture {i}: Width={image_size[0]}, Height={image_size[1]}")

            # Wait for the specified interval before taking the next picture
            time.sleep(interval_seconds)

        # Stop the preview after capturing all pictures
        camera.stop_preview()

if __name__ == "__main__":
    labels = ["label1", "label2", "label3"]  # Add more labels if needed
    num_pictures_per_label = 20
    interval_seconds = 2
    label_delay_seconds = 10  # Delay between capturing pictures for each label
    output_folder = "my_pictures"  # Change this to the desired output folder name

    # Create the output folder if it doesn't exist
    if not os.path.exists(output_folder):
        os.makedirs(output_folder)

    for label in labels:
        print(f"Taking pictures for label: {label}")
        take_pictures(label, num_pictures_per_label, interval_seconds, output_folder)
        time.sleep(label_delay_seconds)

This is how will look like:

After that we now have a new data set that now we can upload our data to Roboflow. Now i’ll talk about the problem with the Raspberry Pi OS. What I tried is installing Roboflow in the Buster version and it didn’t work, it looks like there is no compatible version. Because of that now we need to setup the Raspberry with the latest version of Raspberry Pi OS Bullseye, the problem with Bullseye is that it has a new camera library named picamera2 so this new library is a little bit tricky and the OS also it doesn’t allow you to access the camera easy as in the buster version so is necessary to go to the terminal and activate it manually. To access to this configuration we need to type sudo raspi-config. Warning this doesn’t work with VNC you need a monitor connected to the Pi.

Select the option 3 Interface Options. And will appear:

Then select I1 Legaby Camera enable and then Enable after this will ask you about rebooting and to have the changes made is necessary to reboot it. Then is possible to run the code is really similar to the one in Buster.

import time
from picamera2 import Picamera2, Preview
from PIL import Image
import os

picam2 = Picamera2()

def take_pictures(label, num_pictures, interval_seconds, output_folder):
    folder_path = os.path.join(output_folder, label)
    if not os.path.exists(folder_path):
        os.makedirs(folder_path)

    with picam2 as camera:
        # Start the preview
        camera.start(show_preview=True)

        for i in range(1, num_pictures + 1):
            # Generate the file name for the picture
            file_name = os.path.join(folder_path, f"{label}_picture{i}.jpg")

            # Capture the picture
            camera.capture_file(file_name)

            # Wait for the specified interval before taking the next picture
            time.sleep(interval_seconds)

        # Stop the preview after capturing all pictures
        camera.stop_preview()

if __name__ == "__main__":
    labels = ["label1", "label2", "label3"]  # Add more labels if needed
    num_pictures_per_label = 20
    interval_seconds = 2
    label_delay_seconds = 10  # Delay between capturing pictures for each label
    output_folder = "my_pictures"  # Change this to the desired output folder name

    # Create the output folder if it doesn't exist
    if not os.path.exists(output_folder):
        os.makedirs(output_folder)

    for label in labels:
        print(f"Taking pictures for label: {label}")
        take_pictures(label, num_pictures_per_label, interval_seconds, output_folder)
        time.sleep(label_delay_seconds)

Now after running any of the two previous codes to collect data let’s check how to upload this data from Raspberry Pi to Roboflow using the API.

Upload images to Roboflow using the API and Python.

First we need to have our Raspberry with the Bullseye version check my tutorial to setup the Raspberry. Then we need to run in the terminal

pip install roboflow

If we had success installing the library, open Thonny the Python IDE from Raspberry and click Run with the next code to ulpload the images in the folder you create recently (Just change the corresponding values of your project).

import os
from roboflow import Roboflow

# Initialize the Roboflow object with your API key
rf = Roboflow(api_key="xxxxxxxxxxxxxxxxx")

# Retrieve your current workspace and project name
print(rf.workspace())

# Specify the project for upload
project = rf.workspace("xxxxxxxx").project("xxxxxxxxx")

# Folder path containing all the images
folder_path = "/content/images"  # Update this to your folder path

# Get a list of all image files in the folder
image_files = [os.path.join(folder_path, file) for file in os.listdir(folder_path) if file.lower().endswith(('.jpg', '.jpeg', '.png'))]

# Upload each image to your project
for image_file in image_files:
    project.upload(image_file)

In the Shell will appear all the information and if is running add some print to warning you when it’s over. When is running you will see that it takes a while, but the images will be uploaded.

After it finishes you can wait a few minutes to see the images in the assign section from your Rboflow project as we see in the image below.

Congratulations now you have a new data set of images to label manually and then retrain your model! By harnessing the potential of Active Learning with Raspberry Pi and Roboflow API, you’ve created a streamlined process for capturing images and utilizing the Roboflow label them seamlessly. So, go ahead, dive into the labeling process, and let the possibilities of Active Learning with Raspberry Pi and Roboflow API propel your machine learning journey forward. Happy labeling and happy coding!

Keep an eye I added one extra part How to Deploy a Roboflow (YOLOv8) Model to a Raspberry Pi. I recommend you in case you want to use the full app from part 4 in Raspberry Pi the part 9.If you find errors following this or feedback about this guide let me know in the comments, thank you for following this post. Good luck with your projects.