Community
Select to view content in your preferred language

Smart Assistant Model Training

635
1
07-25-2023 11:01 AM
beachcamp
by
Occasional Contributor
‎07-25-2023 11:01 AM

Hello,

I'm interested in utilizing smart assistants for a survey I am developing. The application is pretty specific, but I think well suited. I want to count specific objects in an image, I'm looking at smart attributes functionality.

I am preparing to train and test a model, I understand it is best to have a large body of training imagery, but I have a question about the type of training images.

Is it better in training to have very focused, subject only images? Or more realistic images?

What I mean by that is if I wanted to train a model to identify stop signs from a picture of a busy street would it be best to train on images of busy streets, or focused images of the target (the stop sign).

Additionally would it be possible to transfer learning from a model like "Common Object Detection" to add additional objects? I'm not sure if that would be a valuable path to take since the images I will be using are quite specific and unique.

Appreciate any help or advice.

0 Kudos
1 Reply
BoruiZhu
by Esri Contributor
Esri Contributor
‎07-31-2023 06:38 PM

Hi @beachcampbeachcamp,

Thank you for interesting in smart assistants! Below are some insights regarding your use case, and I hope you find them helpful.

  • Type of training images: More realistic images are better. In your case, if you utilize your object detection model on images that closely resemble the training images (similar background content, lighting condition, shooting angle, etc.) it would be a perfect use case and example of how you want to collect relevant training images. 
  • It is indeed possible to retrain or transfer a model to include an additional class, but for this purpose, you will require the training file/log rather than just the model itself. The Common Object Detection model in the Living Atlas does not contain the training file, and I don't think it's provided in its original source. And to achieve optimal accuracy and performance, it's better to use your own training images.

Feel free to ask any additional questions you may have without hesitation!

Best,

Borui

0 Kudos