Success with object recognition in stills but need video

[Tom Tolleson]

Hello,

I've trained an ML model in Google Vision on image sets I provided (11 different species of fish). It is is fairly accurate at recognizing fish species in tests on new images.

However, the team I'm collaborating with would like to have these same fish species identified in videos of fish rather than still images. They have supplied me with recorded videos (terabytes worth) and a URL for live video. Either live or recorded or both is fine to work with.

My question is this  - Is it possible to apply my existing still image model to video?
How would I do that? I'm just learning the Google Vision API and not sure how to proceed.

Any advice is helpful thanks!

Tom

[David Mesa Lopez]

Hello, 

I have been researching quite a long time regarding your concern and I am not able to determine any path to follow that will resolve this situation for sure. 

The main issue that I am detecting here is that as my understanding Google Vision is intended for image machine learning procedures and Im not sure if you will be able to use models trained within Google Vision environment to predict videos features in a different Product environment.

For your new scenario use case I would use Cloud Video Intelligence API. In this API you have the opportunity to make predictions with project's models. I would try to follow [1] but using the model ID of the Google Vision trainings.

Hope this information helps you, please share your feedback with us.

Kind Regards

----

[1] - https://cloud.google.com/video-intelligence/automl/docs/quickstart-protocol?authuser=1#make_a_prediction