search

keras-team / keras-cv

Industry-strength Computer Vision workflows with Keras
Other
1.01k stars 330 forks source link

Supporting OCR models #1114

Closed tanzhenyu closed 1 year ago

tanzhenyu commented 1 year ago

We'd very much like to get OCR as part of the offering, preferrably some classical model such as EAST / Rossetta

tanzhenyu commented 1 year ago

@faustomorales Would you be interested?

innat commented 1 year ago

@tanzhenyu It would be great to have more details regarding OCR models from keras-team. Regarding

What sorts of models would be priorities? (For example, some SOTA models: MaskTextSpotter-V3, ABCNet, etc,).

bhack commented 1 year ago

What about https://github.com/tensorflow/models/tree/master/official%2Fprojects%2Funified_detector

tanzhenyu commented 1 year ago

What about https://github.com/tensorflow/models/tree/master/official%2Fprojects%2Funified_detector

Oh I wasn't aware of it. @Jyouhou wdyt?

bhack commented 1 year ago

Note It is unified detector + layout.

It could be ok if I suppose that we want to have a two stage API design here (detection and recognition)

tanzhenyu commented 1 year ago

Note It is unified detector + layout.

It could be ok if I suppose that we want to have a two stage API design here (detection and recognition)

My take here is we would need to have some two stage design anyway (just like OD)

bhack commented 1 year ago

two stage design anyway (just like OD)

Here we could have singe(?) double and triple stage (detector + layout + recognition).

Instead I think that in OD we are 2 stage or 1 stage (e.g. anchor free).

ianstenbit commented 1 year ago

OCR is still something we'd like to include in the long-term but it's not currently a priority