Creating an image dataset is not as difficult or expensive as you might think! In this article, we will be creating one of the most popular datasets for computer vision applications: the MS-COCO dataset. If you are new to machine learning, then the MS-COCCo dataset is a good place to start because it is simple and easy to use.

MS-COCO was originally designed for object detection applications, but it can easily be adapted to other tasks such as segmentation, captioning, and more! The important thing to remember about MS-COCO is that it does not contain any natural looking images. All of the pictures in the dataset have been digitally altered to include specific objects or features.

The main goal of this article is to show you how to quickly create your own image dataset by using free and paid picture sources online.

Take pictures of objects

how to (quickly) build a deep learning image dataset

The next step in creating your image dataset is to take some pictures! You do not need professional photography equipment to get great images. In fact, you do not even have to use a camera!

Any device with an integrated digital signal processor (or computer that has software installed for deep learning) can be used to take photographs. Your smartphone will almost always contain such a chip so it is perfect choice if you are looking to reduce cost.

There are many free picture taking apps available online as well as mobile applications. Some of the most popular include Google Photos, Instagram, and Flickr. Many of these offer users the chance to edit and organize their photos after they have taken them.

Take pictures of people

Creating an image dataset that is already populated with images takes time, resources, and effort. Thankfully, you can easily gather lots of pictures of almost anything!

There are many free sites and apps that allow you to upload and edit photos. Some even have features that let you create new images by adding textures or patterns to existing ones.

By using these tools, creating your own image dataset is easy. In this article, we will be gathering still life photos — those with several objects in one frame.

Why still lifes?

Most computer vision applications require large amounts of data for training. This includes tasks like object recognition or segmenting different parts of an object.

Having enough data is key to ensuring successful learning. That’s why it’s important to start off with as much pre-existing data as possible!

Luckily, there are plenty of free sources of still life photography out there. By picking any item in the photo, you can find what types of items it is under and get some examples of them.

Take videos of everything

how to (quickly) build a deep learning image dataset

There are many ways to take still images, but filming is an even better way to gather image data. You can use your smartphone as a camera or buy a cheap digital camera that you can upload onto a website like Amazon Kinesis where you can have it automatically process and save all of your photos.

Using a webcam allows you to capture more detail than with a normal camera since there’s higher quality source material. By adding some music or other sounds, people create natural settings for you to include in your dataset.

There are also several apps and software programs that will allow you to add captions or keywords to each photo which help identify what kind of item they are.

Take videos of objects

how to (quickly) build a deep learning image dataset

Another way to make your image dataset is by taking video or still pictures of different things. This can be done in nature, at home, or anywhere with a camera!

By filming different items, you’ll get lots of information. The item’s position, color, shape, etc. are all part of making identification easy.

There are many free sources available for you to gather content from. You can use these to make your own small image datasets easily!

Google has a collection it calls Google Cloud Platform where you can upload videos and have computer software analyze them. There is a paid and free option depending on how much data you want to add.

Amazon also allows you to create an account and access their computer vision service that does the same thing. They call it Amazon Rekognition and it costs around $300 per year per user.

But most people will go beyond just having computers identify what something is. They will also look at how it matches up to other similar things so they know more about it. These programs are called deep learning algorithms and there are several types.

Take videos of people

how to (quickly) build a deep learning image dataset

A popular way to build an image dataset is to take video or still pictures of various things and then edit the images in software such as Photoshop, Figma, or other graphic design programs.

By taking these new photographs or videos, you can add all sorts of features that would not be present in the original material.

For instance, if you took a picture of someone at a party, it probably does not have very interesting decorations or clothing. If you filmed the person during the night, perhaps they are lying down and sleeping now.

Take screenshots of everything

how to (quickly) build a deep learning image dataset

A very popular way to build an image dataset is by taking as many, if not all, full-length screenshot videos or still images you can get your hands on and then sifting through those pictures and labels to create your database.

A common misconception about this approach is that it must be done at night, when there are less people around. This assumption comes from the fact that most of us use social media in our daily lives and thus gather a lot of information during the day.

But creating an image dataset doesn’t necessarily have to be limited to only nighttime hours. In fact, one of the main reasons why some of the more famous datasets like Imagenet and MscocoGnet exist is because they were built over several months and years!

By using our own, personal accounts we could easily compile a large amount of data too.

Take screenshots of objects

how to (quickly) build a deep learning image dataset

The next way to make your own image dataset is to take some pictures or videos of things. You can do this by taking general photos or filming yourself doing specific things, like eating, putting products in a cart, or talking.

Alternatively, you could search online for “make me rich quick” or “how to be famous on youtube” then edit the settings and recording software so that you get a lot of screen shots and/or video clips.

These types of images are referred to as screenshot or clip datasets because they contain only photographs or videos of particular items or events.

The most difficult part about creating an image dataset using these methods is deciding how many examples to use per category.

Take screenshots of people

how to (quickly) build a deep learning image dataset

Recent developments in deep learning have given rise to new applications that require large datasets for training. A common application is creating computer vision systems that can identify all sorts of things, such as animals or fruits and vegetables.

A major challenge in developing these systems is finding enough examples of what you want to classify. For example, if your system must determine whether an image contains a cat, then it needs lots of pictures of cats!

If you’re already taking good pictures, then there are many free ways to gather images with apps and sites across mobile devices and laptops.

But what about those who don’t? Or those who want to quickly create their own image dataset instead of buying one?

Well, you now have a way to do it! By our definition, a “screenshot” is just part of a picture that includes some text. So, we’ll be gathering image + caption data by taking random screenshots online and labeling them.

Caroline Shaw is a blogger and social media manager. She enjoys blogging about current events, lifehacks, and her experiences as a millennial working in New York.