How to train Excire 2.0 to add keywords

[Peter Gallagher]

I updated to 2.0 but am disappointed by it's keywording performance. I don't need elaborate keywords (one or two usually) but Excire 2.0 seems to miss the obvious. My LR catalog has several dozen images of paintings from galleries, usually in a frame on the gallery wall. None of them is anything but a plain representation of the actual painting. When I re-initialized my database for Excire 2.0 it attached the "Painting" keyword to just 5 of all of these images. There is no obvious reason why.

In fact, several of the images that Excire 2.0 failed to classify with the "Painting" keyword in its database had a keyword 'painting' in the keyword field of my LR database!

I want to re-train the search engine to attribute the Painting keyword to the other paintings. A couple of years ago (or more) you indicated on this forum that you would provide re-training options for Excire. But nothing seems to have been done about this.

How can I remedy this deficiency in my Excire 2.0 database? I suspect I'll find other similar examples of wrong or deficient keyword attribution.

[ExcireThomas]

Hello Peter,

thanks for your suggestion, but from an algorithmic point of view this is not possible. Please notice, we trained our system with more than three million images. For each class we have at least 1000 samples, for most classes we have usually much more examples. Furthermore you need special hardware which needs several days for training. So a client side training is not possible. This is how AI works nowadays.

But we plan to develop an online portal, where our customers can provide their photos and label information for training and refinement of our software, which we improve continously.

Best regards,

Thomas