Does anybody have first hand #DigitalHumanities experience with SAM3?
The web demo is crazy good, also with historical data but we found that performance unfortunately doesn’t really translate to actually using the model (had similar experiences with earlier versions of it).
At the same time, I’m a fan of SAM so it would be great to be able to get it live up to its promise 😅
@felwert @sarahalang we've been using SAM2 in https://immarkus.xmarkus.org. Worked ok for us. SAM3 is something I'd like to try. But project didn't have the time/resources yet.
Overall, I guess it depends on:
i) your expectations :-)
ii) the material you're working with (obviously)
iii) whether you're using the prompted or unprompted mode (we're only using prompted in IMMARKUS)
iv) most importantly, the model size/variant
In our case, we used the "tiny" variant only, because that fits entirely into the browser – no server-side embedding needed. The larger variants would give better results I'm sure. (But then you'd need a separate server component which you need to maintain, fund, keep running... not DH-friendly as we know ;-)
An image annotation environment for the MARKUS platform. Developed by Prof. Dr. Hilde De Weerdt, Dr. Rainer Simon, Dr. Lee Sunkyu, Dr. Iva Stojević, Meret Meister, and Xi Wangzhi with funding from the European Research Council under the Horizon 2020 programme, grant agreement No. 101019509.
Sarah Lang (she/they)
Frederik Elwert
Rainer Simon