[anh_n]

Hi all, I am co-founder of Lookuq and want to explain a bit more about our idea.

There is a lot of apps to recognize some things. Popular example are PlantNet, PlantSnap for recognizing plants. They all have similar functionality, you take a picture of the object and then the app show you a list of objects it thinks what is it. Click on it, you can see more information about the object.

There are many things which visual identification is helpful, e.g. dog breed, animal, plant species, vintage watches, artworks, ...

Everyone is expert in something and may have a lot of images of the objects and information about it. They should be able to share this knowledge to other people.

With our Lookuq platform, it is super easy to create an object recognizer app without coding. All you need is photos of the objects and some information about the objects in tabular form (CSV, Excel format). Then you can upload it to our website, let the system automatically create and train the model to recognize the objects. After that just use our pre-made app (available on both iOS and Android) to recognize the objects and show related information to the objects.

I would love to know what you think about it?

[zanew101]

Maybe this is something you want to keep internal to your company, but can you say anything about the model architecture being used here?

[anh_n]

I can not tell much about the model architecture. We have some proprietary technology to reduce the amount of images needed for training and to increase accuracy.

[whoisjuan]

- Will you ever have an API for this?

- Will this work with abstract elements like symbols rendered on a screen?

[anh_n]

Yes, there will be an API to use the object recognizer with other apps or website. We have pre-made both the iOS and Android app because many users can not or do not have time to create the app.

And yes it will also work with any symbols or diagrams rendered on screen. Almost everything we human can visually recognize will work.

[thedjinn]

How many images does your app typically need to avoid overfitting?

[anh_n]

The number of images depends on several factors like the total number of objects, the variation of images, visually difference between objects, etc.

We have developed some proprietary technology to reduce the amount of images needed. In general each object should have 10-100 images to avoid overfit.