Are You Ready for Our Next-Gen Visual Search Engine?
We hope so! It won’t take much longer until pixolution flow 4 will be released. With this major release everything is going to change: the underlying architecture of our product as well as its scope of functionalities. Curious what benefits are in for you? Let’s take a look at some highlights we’ve packed into this release preview.
pixolution flow 4
We already shared some details about pixolution flow 4 in the first part of our release preview which focused on new features and new capabilities thanks to a complete redesign of the architecture.
Today, we dive into the improvements of our core functionalities which are going to be better than ever thanks to our new pixolution image descriptor. With the support for multiple images per document we introduce some awesome new application possibilities. Last but not least, we summarize how pixolution flow 4 can be operated in a fast moving DevOps environment.
Improved Core Functionalities
We made extensive efforts in R&D and it was all worth it: we finished our new pixolution image descriptor. As the name implies, the job of the descriptor is to describe the content of an image. The new descriptor is the basis and the heart of pixolution flow 4 and it’s designed to work with any conceivable image sets out of the box.
We have done intense testing with many different image sets. The result
is an increase in mean Average Precision of 11.6 percentage points compared to the previous image descriptor. While this may sound little exciting that’s a massive improvement you will recognize!
What does that mean for you when using pixolution flow 4? The quality of query results will be tremendously better. Searching for visually similar images in your image database will return way better results. Detecting duplicates will be more robust and gives you more solid results. If your need is to tag images, you’ll be delighted to hear that the tag suggestions will be more precise.
And yet, it could be the case that our improved image descriptor may not be good enough for you, which is when you have a special image collection. Let’s say you are a screw manufacturer and you have product images of all varieties of screws. Our general purpose image descriptor may not be able to distuingish these images sufficiently. Instead, you would need a specialized image descriptor capable of recognizing the subtle differences between screws. Fortunately, that’s not a problem. Thanks to our new architecture in pixolution flow 4 we can develop a custom AI module for your pixolution flow instance which is optimized for handling your special image collection.
New Application Possibilities
With pixolution flow 4 at hand you’ll not only have way better query results, you can also benefit from completely new application possibilities. pixolution flow 4 will be able to index multiple images per document. Here are some example use cases:
1. Search for product variants
Products in online shops are mostly represented by several product shots. They show the product from different angles, with different colors, worn by a model and so on. pixolution flow 4 can index all of these images as belonging to one product and then take take them into account when doing visual searches.
2. Close-up search
Often, an image contains more than just one object and it is not clear what is actually the important part. It might be as small as a ring, although it takes so little space. With pixolution flow 4 you can adress those scenarios and index close-up images as well. If a user searches for specific ring, pixolution flow 4 will find one of its close-ups and you can show the associated main image. The user experience will be great.
3. Index 360° images
If you have 360° panorama images you can split the image into single sections and index them as well. The search quality will be way better, since pixolution flow 4 is then able to find specific objects and the approximate location in the 360° image.
4. Search in videos
The same goes for videos. Instead of indexing just one master keyframe, you can index several frames. This way it’s possible to find specific scenes or objects in videos. Since pixolution flow 4 can store metadata for each image, you can also index the timecode to jump straight to the correct position in the video.
Easier Operations When Analyzing and Indexing
Finally a few words about different operating scenarios. Analyzing images is a time consuming process — compared to analyzing text. To avoid putting heavy load on your search server, you can split the analyzing and indexing processes into two steps.
pixolution flow provides APIs to analyze an image and return the image descriptors as JSON text. This text can be stored in any arbitrary database alongside your other existing data. You then just index all of your data (including the preprocessed image descriptors) into pixolution flow 4.
Decoupling the analysis and indexing processes brings you several operational benefits:
- Load is taken away from your search server.
- Faster analysis: Offline computation can be done on dedicated hardware (like GPU servers).
- Fast indexing since pixolution flow only validates the precomputed data.
- Fast reindexing: Analysis has to be done only once and not every time you have to reindex your data (because of software updates or data changes).
We hope you are looking forward to the official release of pixolution flow 4 like we do. Please have a little more patience with us, we’re about to do the finishing touches.
You want to know how pixolution flow could help solving your specific challenges with visual content? Drop us a line. We’d love to hear from you!
Originally published on https://pixolution.org/blog.
