Getting started with Roboflow

Roboflow is a platform to help developers and researchers build, manage, and deploy computer vision models. It simplifies the process of working with image data for machine learning by providing tools for data annotation, data augmentation, model training and model deployment.

Step 1: Create an Account on Roboflow

Start by creating an account on Roboflow using your NCSU email ID.

Sign up for the research plan

The research plan increases the quota of training credits and other training limits for students. To activate this, go to Settings->Plan & Billing->Research plan.

Note: You must use a NCSU email address for activating the research plan.

Step 2: Create a New Project

Start a project by clicking on ‘+ New Project’

Open

Enter a project name and select the type of task you want to perform. There are various types of computer vision tasks based on the goal of a project. Learn more about the types of computer vision tasks: https://www.v7labs.com/blog/what-is-computer-vision#7-common-computer-vision-tasks (For the hackathon the task is object detection).

Click on 'Create Public Project' to proceed.

Step 3: Upload Dataset

Upload the images from your dataset. Optionally, assign a batch name. Roboflow supports data in the form of both images and videos, extracting frames at the rate.

We can also add annotation files for each image. For the hackathon, we will skip this step as we will annotate the images ourselves.

Click ‘Save and Continue’.

Step 4: Annotate the Dataset

Roboflow supports both Auto Labeling (limited by subscription) and Manual Labeling.

Click on ‘Start Manual Labeling’ to start annotating the data.

Teammates can be invited to your project to collaborate and the images can be divided among the teammates to annotate the images efficiently.

Start drawing bounding boxes around the objects in the image. There are multiple tools to help draw the bounding boxes. After drawing a bounding box, you need to select the class which it belongs to.

The summary of the current progress is shown here. After annotating all the images, click ‘Add x images to Dataset’ button to save the annotated dataset. This annotated dataset can now be used to train computer vision models.

Step 5: Prepare the Dataset for training

Set the train, validation and test split according to your preference. Here's a guide to help understand these splits:

Apply any desired preprocessing and augmentation steps to the dataset (optional). Learn more about importance of preprocessing and augmentation here:

https://medium.com/@saba99/preprocessing-techniques-182c8ac186f6

https://aws.amazon.com/what-is/data-augmentation/

After configuration, click ‘Create Dataset’ to generate the version of your dataset. You can always create different versions using various splits, pre-processing and augmentation techniques.

Step 6: Train a Model

Once the dataset is ready, click on ‘Train with Roboflow’ to start training a model.

Select the model type to train the dataset on and click ‘Start Training’. Explore different methods of training including training from public checkpoint, random initialization and explore the differences between them.

After training is complete, review the model performance metrics. Adjust the parameters and use different techniques to train different models to improve the evaluation scores.