Introduction

AI/ML application has taken off in recent years. This has mainly taken place in the Enterprise/Corporate world or at specialist startups/scaleups. The general and broad SME has so far invested limitedly in the possibilities that AI/ML can offer, e.g. in the context of the digital transformation. In the past 2 years, various initiatives have been started to change this. We regularly work as an R&D partner for smart products and services with SMEs. AI/ML is playing an increasingly prominent role in this. That is why the question has arisen whether there are already AI / ML software platforms that are effectively usable for SMEs.

Last month, we therefore conducted a market survey into an integrated AI/ML platform. The central question for this market orientation was: is there an AI/ML software platform suitable for SMEs on the market, for rapid application development of Computer Vision applications?

This report presents the results of that market orientation that we have carried out. The result can be used for your own (further) orientation and/or selection. Or to get an idea of our competence and to engage us for a case that you would like to have assessed with one of the solutions listed below, or another.

A market survey has been made that is based on previous insights and experiences. About 4 years ago we conducted a similar study, which was supplemented with our own new insights and professional developments. At that time we participated in the IBM Startup program and compared IBM Watson with a number of other players.

In the overview below, vendors such as Microsoft Azure ML and AWS are also missing. With this exploration, I don’t pretend to be complete. We have made a pre-selection of vendors that are closer to us, including from the earlier exploration.

Features of AI/ML development platform

It is an end2end platform for the development of ML models and applications with images and video as input and focused on the general and also broad Computer Vision scope. As a result, various other application areas and media types have also been explicitly excluded, such as: text, sound, speech, etc.

For the chosen betting option, the following core functionality is necessary:

• Data preparation
• Labeling of data manually and/or (semi) automated
• Model learning
• Model deployment
  • Application integration both Cloud based (API)
  • Edge Computing for local real time inference and Internet independence during run time

Ideally, the following properties are also present:

• The (business) job goal must be able to determine results instead of specialist AI /ML knowledge of all the different ML models that may be suitable for the job goal
  • See the image below as an example of function goals

• API, Python preference and CPU and NVIDIA GPU hardware support
  • with Intel OpenVINO as a possible additional alternative
• AutoML has a certain preference with the possibility to create custom models at a later stage to enable wider use than CV alone!

Why prefer AutoML?

AutoML is one of the more recent developments in ML. AutoML makes an important contribution to the desired characteristic of this market orientation:

“The (business) function goal must be able to determine results instead of specialist AI/ML knowledge of all the different ML models that may be suitable for the job goal”

On the Web-site https://www.automl.org, you can read this about it:

“AutoML …

… provides methods and processes to

• make machine learning more accessible

• improve efficiency of machine learning systems

• accelerate research and AI application development

Machine learning (ML) has achieved considerable successes in recent years and an ever-growing number of disciplines rely on it. However, this success crucially relies on human machine learning experts to perform manual tasks. As the complexity of these tasks is often beyond non-ML-experts, the rapid growth of machine learning applications has created a demand for off-the-shelf machine learning methods that can be used easily and without expert knowledge. We call the resulting research area that targets progressive automation of machine learning AutoML.”

All this fits beautifully with the questioning of this market orientation. And as it turns out later, 2 solutions have been considered, which have embraced the concept of AutoML.

Method of market orientation

The following method has been applied.

• Orientation and preference selection for a later trial implementation in the form of a PoC for which there is a maximum budget of 100 hours from an experienced business /ICT consultant
• By applying it, gain practical experience and more knowledge
• Evaluate and, if necessary, revise the previously made preference

During this market orientation, a concise trial application was developed for the platforms considered that did not have a shoot out on pricing (see also below). So we not only looked at specifications on a Web-site, but also actually assessed a practical result.

Overall preference criteria

• 3 main criteria
  1. Affordable for SMEs to start with development and deployment (price level max. €250 / month)
  2. Is the suitability of the solution sufficient?
  3. Does the supplier show sufficient confidence to offer immediate application value, but also to continue to invest in relationship and technology as an SME?
• Without a published and clear pricing, there is a shoot out of the vendor/service
• This setup is based on our experience with product/market combination for SMEs. In our experience, all vendors without published pricing focus directly or indirectly on the Enterprise/Corporate segment. This makes them not the most interesting for SMEs.
• The consideration – see below – of all products/suppliers stopped immediately after the first criterion; after all, it is taken as a shoot out!

Overview of platforms considered

We only looked at commercial providers, despite the fact that there are also interesting (commercial) Open Source products such as auto.gluon.ai. The reason for this is the specific SME target group that, with little knowledge of the matter, still wants to successfully apply AI/ML in a limited time and with limited AI/ML expertise. In my opinion, the Open Source platforms are not suitable for this at the moment.

It is deliberately not talked about assessed,but considered. To make a complete and balanced assessment, more time and knowledge is needed. Which was not available in the limited set-up of about 60 hours of this activity.

• cnvrg.io
• Google Vertex AI
• IBM Watson
• Chooch.ai
• Always.ai
• Viso.ai
• Clarifai
• Roboflow
• Nyckel Hotels
• Edge Impulse

Not so much as end2end AI/ML platforms, but as additional and/or partial solutions, we also looked at:

• NVIDIA Triton Inference Server
• v7labs
• Superannotate
• Labelbox

These are not taken into account; please contact us if you want to know more about it.

Finally, it is good to report that the classic Machnie Vision suppliers such as MVtec with Halcon, Adaptive Vision and Matrox also offer AI / ML modules. These are specialized and are not yet understood as an end2end AI/ML platform with predefined properties. That is why they have been disregarded.

end2end AI/ML platforms considerations

Only the vendors/products marked in green meet (in full or in full) the set criteria. The overall result of the consideration in order of preference:

1. Google Vertex AI
2. Clarifai
3. Nyckel Hotels

The speed at which a trial application could be developed (an Image classifier with its own trained model), looks like this with the fastest first:

1. Nyckel Hotels
2. Google Vertex AI (in AutoML mode)
3. Clarifai

With the preferred criteria given:

1. Affordable for SMEs to start with development and deployment (price level max. €250 / month)
2. Is the suitability of the solution sufficient?
3. Does the supplier show sufficient confidence to offer immediate application value, but also to continue to invest in relationship and technology as an SME?

has become the final ranking of preference for our application:

1. Google Vertex AI,with caveats due to the somewhat complex pricing and yet to be investigated optional addition of Labelbox and NVIDIA Triton Server for Edge Computing optimization
2. Clarifai,does not yet have its Edge Computing part generally available but otherwise it partly meets the ideal profile and is therefore further followed
3. Nyckel,a Y-Combinator startup for AutoML as a Service, simple, powerful, affordable, just now no Edge Computing possible, with uncertainty about the existence in the solid force field of large(er) AI/ML suppliers.

Finally

Our first AI/ML application for SMEs is being developed using.b Google Vertex AI. In addition, we will investigate what value Labelbox offers for data preparation and how we can realize Edge Computing ourselves by exporting Tensor Flow saved model. More on that in a few months.

This report presents the results of a market orientation we conducted in December 2021. Please contact us if you want to know more about it or want to discuss an AI/ML issue with us.