We are always interested in to see what’s going on around us. Recently we have prepared an extensive image recognition test and performed it on our API and two of our major competitors: IQEngines and Kooaba. The test set consisted of hundreds of real-life pictures of books, CD’s, posters and other such products. We have gathered tons of statistics and were delighted to see just how great our technology performs in this context. This article presents extract from these results.
First matter we have taken into consideration is the time required to create an account. In case of Recognize.im it’s about 30 seconds. It is fully automatic and dead simple, just like it should be. However things didn’t look so great in case of our competitors. IQEngines barely made it within 24 hours and Kooaba… well, they didn’t meet the declared 48 hour deadline. When someone thinks of a long-term partnership this probably doesn’t make that much of a difference, but before that happens, someone might probably want to try a quick demo and that’s what our rivals simply fail to provide.
But this was just the beginning. Once we have gotten our accounts ready we
started to feed them with reference images. Each platform performs this in a different way and I will not go into details here.. For your convenience and simplicity we have measured the total time it took from the moment we started to upload the first reference image until all of our reference images were uploaded and ready to be recognized. We have tested different sizes of image sets. With small sets these preparation times were fairly equal, but once the numbers reached thousands our vantage became apparent. Here’s the summary for 1000 reference images:
As you can see not only did our solution turn out to be the fastest, but also the most consistent. From the two other solutions particularly IQEngines showed worrying values. Sometimes they were able to set up the system in 30 minutes, which was comparable to our performance, but sometimes it took them almost three times longer than us. This is very unpleasant feature because it means that you will never know just how long certain processes will last. Our solution is very reliable in this regard, and this, in practice, makes a big difference.
When all three accounts were fueled up and ready, we proceeded to our main benchmark: the recognition. Hundreds of pictures, one by one, where sent to each API and for each of them recognition time and correctness of the response were recorded. The statistics exhibit a very clear picture of the situation. Recognition rate of all three participants was very similar, but recognition times were dramatically different. This held true regardless of the size of reference image set. Please consider following chart:
This clearly shows our most significant advantage over competition: speed. Recognize.im is as effective in recognition tasks as any other state-of-the-art solution, but works many times faster. Our superiority is further enforced by the fact, that once again our solution turned out to be the most reliable in terms of repeatability. This is also reflected in the worst case scenarios. With Recognize.im the longest recognition took less than 3 seconds, while IQEngines and Kooaba were digesting theirs for 62 and 21 seconds respectively.
This is a major achievement that has some profound implications. First of all high speed means great user experience – in most cases recognition is done instantly without users even knowing that they were waiting for a response from a server. In very rare cases it will take few seconds, which is still acceptable. But that’s not all. Extra short processing times allow for a whole new range of applications such as real-time scanners with augmented reality. With recognition rate of 3 frames per second a lot of really cool things can be done. In the near future we plan to show you some examples of such use cases… but let’s not spoil the surprise.
We have tried to keep our benchmark as objective as possible, but, however we’ve tried, you can always argue that it was more or less biased in our favor. There is really no way for us to convince the unbelievers, therefore we encourage everyone to perform their own comparisons and find out for themselves. After all it is free and takes only 30 seconds, right? Oh, wait, that’s just us. You have to wait a few days for the other two.