Divar’s AI: A Catchy Slogan or Practical Reality?
For a while, I had been hearing about Divar’s new service called Retouch, and today I decided to check it out to see if it was as practical and usable as claimed or just another marketing gimmick. But before that, let’s first examine the technologies that services like Divar’s use.
Convolutional Neural Networks (CNNs)
Convolutional Neural Networks (CNNs) are among the most widely used technologies in image processing and object recognition. CNNs can identify and remove objects by analyzing image features at different levels.
Generative Adversarial Networks (GANs)
Generative Adversarial Networks (GANs) involve two neural networks: a generator and a discriminator. The generator attempts to create more realistic images, while the discriminator tries to differentiate between real and generated images. This process enables the model to produce and accurately reconstruct removed parts.
Inpainting
Inpainting is a technique used to fill in empty or damaged areas in images. AI models can use existing image data to realistically fill in these gaps. This technique is particularly useful for object removal and reconstruction of deleted parts.
Deep Image Prior
Deep Image Prior is a new technique that uses the structure of neural networks to reconstruct images. This method does not require external training data and relies solely on the information present in the image to restore removed sections.
Contextual Attention
This technique utilizes the Attention Mechanism to identify relevant contexts and fill in missing parts. AI models can focus on surrounding textures to more accurately fill in the gaps.
The service is aptly named “Retouch,” considering its functionality. It is provided by a startup called Retouch in collaboration with Divar. Now we know that it employs a combination of the aforementioned technologies to perform its tasks.
Since the service is currently in a pilot phase in Mashhad and Isfahan, I traveled to Mashhad and changed my location to activate the service. Here’s how it works: Suppose we have an apartment that we want to list for sale or rent but don’t want our belongings to be visible to the public. The first step is to register the listing, which then goes into a waiting status for review and publication. At this stage, you see two options: “Other Complements” and by selecting them, you’ll find “Retouch” and “Stochy.” I selected Retouch and encountered a 404 error page 😊, likely a web service issue. However, I received a link via SMS that led to the retouched images. The preparation speed was impressive, and I hope that Retouch can handle the user load and the GPU server costs involved.
The results were more impressive than I expected. In the output image (the video of which I’ve included here), you can see the result. If you have experience with many AI image processing tools, you know that image details are always the hardest and most crucial part, and Retouch has managed this well.
One downside I noticed is the lack of object selection capability. For instance, you can’t choose to keep certain objects in the image, and I think this feature should be added to allow more control over the image. Perhaps offering a premium subscription for access to these customization and editing features would be a good idea.
Congratulations to the Divar team for launching such an exciting service! I hope they also address the strange pricing models, like listing prices as 11111, which complicates the search filter issues.
