A new SOTA text recognition architecture - SVIPTR

[milosacimovic]

Hi,

I would like to suggest possibly introducing another state-of-the-art text recognition architecture to docTR.
SVIPTR
It's promising accurate results at low latency.

Notably, the SVIPTR-T (Tiny) variant delivers highly competitive accuracy on par with other lightweight models and achieves SOTA inference speeds. Meanwhile, the SVIPTR-L (Large) attains SOTA accuracy in single-encoder-type models, while maintaining a low parameter count and favorable inference speed.

Thanks for your consideration.

[felixdittrich92]

Hi @milosacimovic 👋🏼,

Thanks for sharing with us, I will have a look on it after vacation 😊
Does not look like there are benchmarks against parseq / vitstr actually right ?

[milosacimovic]

Hi @felixdittrich92 👋,

Happy holidays! 🎄
Here's their benchmark table
It does not include parseq but it seems it's a bit less accurate on IC13 dataset. However, the large variant seems to be better on average than some really big models like TrOCR for example while being a lot faster and smaller.

[felixdittrich92]

Thanks wish you too a few days to relax 👍🏼🤗
I quickly checked the paper and benchmarks today and it looks like a valid candidate to be added to docTR .. so feel free to open an issue ticket.. from inference latency it should be comparable with crnn_mobilenet_v3_large by keeping the capacity comparable to parseq☺️

[felixdittrich92]

Would you like to work on it ? :)

[felixdittrich92]

#1867