Support for Custom Labels in Rekognition

[040medien]

I couldn't find a mention of Custom Labels in the documentation, so I'm guessing this isn't supported at the moment, correct? Do you have any plans of supporting these in the foreseeable future?

Many thanks!

[mgottholsen]

I would love to have some sort of custom label support as well. I've got a use case where I would like to use it for sound identification with spectrograms generated from my security camera network with FFMPEG & the Unifi Protect HA integration. Calling against the custom label dataset from the Urban Sound 8K dataset, converted to spectrogram form as well, it should theoretically work with some wiring :)

[robmarkcole]

I've used custom labels at work, and the issue is you basically need a dedicated EC2 instance running for them, which is much more costly than using just regular recognition. If someone wants to create a PR to support them Im happy but think its quite a niche use

[fberbert]

I've used custom labels at work, and the issue is you basically need a dedicated EC2 instance running for them, which is much more costly than using just regular recognition. If someone wants to create a PR to support them Im happy but think its quite a niche use

Creating custom labels for me is not feasible. I created a model, trained it, and turned it on. The cost is very high. In just 3 days my billing came to around USD 500. When I noticed the cost I canceled the project immediately.

Note: I was able to use Custom Labels in HA using the "node-red-contrib-simple-aws" add-on for Node-Red.

[robmarkcole]

$500! Custom is always going to be expensive in hosted platforms. Alternatively checkout my Deepstack integrations which support custom models on local device