Guide on how to Outsource Image Labelling

What is Image Labelling?

For a better understanding of image labelling and data labelling, let us take a personal example.

Ever since digital technology became popular for taking and storing pictures, instead of the traditional film that could be used only once, it has freed us up to take more and more pictures, especially when combined with the ubiquitous smartphones, one of which is in everyone’s pocket or hands. 

So much so that it is now unclear if the main function of a smartphone is to click pictures and videos, or to communicate through voice calls and messaging.

As we started taking more and more pictures, they also started getting stored and backed up, on physical storage devices or on the cloud.

Either way, many of us have ended with a huge amount of data in the form of pictures that we have clicked over the years.

What happens when we need to retrieve a picture? Presumably occasional retrieval would be the purpose of the photographs.

If we never need to retrieve them, perhaps we should never have taken them.

Say, of a child appearing for the first time in a play in school.

Or, pictures of my grandmother who is now no more. How do we retrieve these pictures?

For most of us, the only way to do it is to go through the data image by image, or file by file. 

For specific events we may still have a chance as we may know the approximate date and could focus on that section of the data.

But for something like grandmother’s pictures, we have to go through the entire set.

Why is that? If taking images and storing them has become smarter, why hasn’t retrieval?

It has, but it needed some effort on our part at the beginning.

Of tagging each image in a way that it would submit itself to later retrieval.

So, while storing, if we had taken the trouble to identify the people in these images and tagged them, it would have enabled us to run a query and retrieve images accordingly.

This, in a way, is Image Labelling. An enrichment of image data. Which is what image labelling outsourcing will do.

Of course, taking it a step further, Image Labelling is a key input to the development of an Artificial Intelligence (AI) engine.

It enables a machine to understand data.

Going ahead with our example, if we identified pictures with grandmother and connected this with attributes of the image that could be understood by a machine, say skin color, or height, or distance between the eyes, at some point of time the machine will begin to understand that if there is a person with these attributes in an image, it is grandmother.

At this point you don’t need to label any more pictures. 

Your AI engine will now be able to identify her even from images where you have not marked her out in advance.

This is the relevance of Image Labelling in AI.

Techniques of Image Labelling

Whether you outsource image labelling or do it internally, there are many techniques involved. A few common ones are described here:

Bounding Boxes

This is perhaps the most common type used for training data sets.

It is easy to visualise as a two-dimensional frame around an object that needs to be called out, like a car, or a house. 

It is placed around an object to define the location and size of the object relative to the entire image. 

Its position can be specified as a set of x and y co-ordinates, typical for a two-dimensional image.

This technique is used for detection of objects and localisation tasks.

Polygonal Segmentation

Most objects are not rectangular. Even a car. 

For example, it is not perfectly rectangular. It can be considered as largely rectangular which is why it can be expressed inside a Bounding Box. However, many objects cannot be. 

An animal for example or a flower.

Instead of identifying an object within the confines of a rectangular figure, polygonal segmentation provides the flexibility of bounding an object within a shape which best fits around it. 

This enables the object to be identified in a precise way.

Cambridge.com defines a polygon as a flat shape with three or more straight sides. A square is a polygon. A triangle is a polygon. An octagon is a polygon.

Semantic Segmentation

In this technique, every pixel in the image is assigned to a class with each class being defined. 

For example, in an image of a street, one might have classes like road, building, sidewalk, cyclist, shop, etc.

This is used where the ambient environment and surroundings are relevant, like when we train self-driving cars. 

They need to be able to pick out a stop light from a pedestrian crossing.

3D Cuboid

This technique could be seen as similar to the Bounding Boxes technique. 

The additional layer of complexity is created by adding depth as the third dimension, making the object a 3D figure. 

This also allows calculation of volume of the object, which may be relevant in some use cases.

Self-driving cars are a commonly cited example these days. If a self-driving car has to wait for an oncoming vehicle to pass before moving, it needs to know not only the rectangular face of the car, but also the depth; how long it is, which will enable us to calculate the wait time before the self-drive can move.

Key-Point and Landmark

This technique is generally used for detection of small objects and shape variations. 

This is done by creating a set of ‘key-points’ across the image, where each combination would have a meaning. 

Points arranged in one way could mean a smiling face while arranged in another way could mean a sad face. 

Another way of defining it would be as marking out features on a landmark.

Should you outsource image labelling?

We are familiar with the GIGO (garbage in, garbage out) principle. The output, mostly, from a business process, will be as good or as bad as the input.

As the output from image labelling outsourcing forms an input into Machine Learning which feeds into the development of an AI engine, it can have long-term implications, hence should be done carefully and with requisite checks and balances in place.

There are many pros and cons to leveraging the image labelling outsourcing offered by leading providers. 

Suffice it to say, and perhaps as a sign of evolution of the industry, the verdict is largely in favor of outsourcing. At the very least, when you outsource image labelling, you can expect the following benefits to accrue:

Engaging a specialist team that has signed up for the task of data entry, including image labelling. 

This team deploys technologies that facilitate the task and enable it to be done efficiently. 

Moreover, the resource cost is typically lower than inhouse processing since the resources hired possess minimum qualifications and are provided on-the-job training to make them fit for purpose.

As a result of a specialised workforce, purpose-built technology and processes and control mechanisms honed over an existence of doing similar work, are likely to deliver work of a better quality.

Leaving your specialist resources to continue to work in jobs that they have knowledge and experience in, and continue to create value for the business in those roles, instead of being bogged down with monotonous, repetitive work that they are not trained for or have ever expressed interest in, and delivering poor results as well as getting demotivated.

Criteria for Selection of Provider for image labelling outsourcing

For better understanding, evaluation of interested providers can be done on two major parameters, which can be further drilled down into multiple capabilities.

The two parameters are:

• Contextual Capability

This is meant to ensure that the provider will be able to deliver the required quality of work once he takes it on. 

The provider should either possess skill and knowledge of the work that they seek to take over, or provide convincing arguments as to why, in the absence of prior skill and knowledge, they will be best placed to provide the greatest value when you outsource image labelling work to them.

• BPO and Organisational Capability

This parameter is meant to be an evaluation of the many enabling services that facilitate delivery by all the units in the organisation responsible for different projects, when you seek to outsource image labelling. 

Without these enabling services, the delivery unit will often not be able to do justice to the client.

Let us dive deeper into the capabilities that make these up.

Contextual Capability

The provider should:

Demonstrate experience of performing similar work with good results – oWorkers has referenceable clients willing to testify to our delivery capability in providing image labelling outsourcing. 

With over eight years of experience over multiple engagements, oWorkers offers unmatched experience, skill, consistency and variety in its services.

Have independent processes that control and deliver quality and accuracy – oWorkers offers industry leading QC (Quality Control) and QA (Quality Analysis) that require a strict regimen of sampling a set of primary transactions performed by agents to ensure they are compliant with guidelines and are delivering the quality and accuracy that oWorkers in known for. 

The sample size is increased or decreased keeping in mind client requirements as well as performance of individual agents and teams. The QA teams also actively coach delivery teams to drive best practices.

Be in a position to offer a wide spectrum of services under the umbrella of Image Labelling Outsourcing – oWorkers has over eight years of experience in providing image labelling outsourcing using different standardised techniques like Bounding boxes, Polygonal Segmentation, Semantic Segmentation, 3D cuboids, LIDAR segmentation, Key-point and Landmarking and Line Annotation.

Offer multilingual capability to cater to global footprint as well as future expansion plans – With three global delivery centers that draw their workforce from the local communities, oWorkers now supports 22 languages for its data entry solutions. 

These language capabilities are available to our clients who outsource image labelling to us, just like all other clients outsourcing other data entry projects to our team.

Sign up to SLAs that offer quick turnaround for delivery – With centers located in the most preferred regions of the world for outsourcing, supported by a continuously operating machinery that runs 24 hours a day 7 days a week, oWorkers provides 24 hour or overnight turnaround on many projects. 

In fact, a quick turnaround for all work is our preferred model, which enables us to do even more for you.

Provide competitive pricing, with flexibility

The location of our centers which are amongst the most favored BPO locations in the world, and access to the local community for drawing resources from, makes oWorkers price competitive for clients from around the world. 

We often enable clients to save upto 80% of their costs when they outsource to us, especially from the US and Western Europe. 

And without having to compromise on quality or timing. When you outsource image labelling to oWorkers, you get automatic access.

BPO and Organisational Capability

These are organisational capabilities that support and augment the delivery effort. The provider should:

Deploy updated technology with data security – oWorkers connects with state-of-the-art technologies through its partnerships with technology owners and providers, and offers them for use of its clients. 

We are ISO (27001 and 9001) certified. Our staff sign and operate under an NDA (non-disclosure agreement). Where required, we offer physical segregation of projects of different clients.

Provide Scalability, and the ability to hire and train resources as required – With its model of working with employed staff, and not contractual or outsourced staff, oWorkers provides flexibility to clients in changing jobs and roles at short notice, as required by a client. 

With deep involvement in local communities, we remain an employer of choice and can offer a scaling up ability of 100 resources within 48 hours, significant for any region of the world.

Display regulatory and financial Stability – Operating as a locally registered company in all three centers in three different geographies of the world, oWorkers is intertwined with local lives. 

It pays local and social taxes on behalf of its staff and complies with all regulatory requirements. The organisation has been a profitable enterprise, with costs carefully controlled and monitored. Operating from the Eurozone, we are necessarily GDPR compliant.

Demonstrate Management Support for the engagement – With a hands-on top management team with many years of experience in running this business, the top team is deeply immersed in delivery, provides direction to the team and often directly interacts with clients

Conclusion

As a pure provider of data entry BPO solutions, with multilingual capability, oWorkers has few peers. 

That is perhaps the reason why several unicorn marketplaces have chosen to work with us.

Your partnership with us for our image labelling outsourcing will bring positive social and economic change through employment in underserved communities, as well as usher motivated individuals into the global digital economy.