• State
  • Locked Locked
  • Replies 2 replies
  • Subscribers 30 subscribers
  • Views 299 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Does TI boards have support for OCR models?

Sourabh Patil
Prodigy 146 points
Other Parts Discussed in Thread: TDA4VM

Hi,

I want to try OCR models on TDA4VM (SDK link: https://software-dl.ti.com/jacinto7/esd/processor-sdk-linux-edgeai/TDA4VM/09_00_00/exports/docs/common/sdk_overview.html)

I could not find any models in model zoo for this (let me know in case I am missing any). I wish to use custom models for this. As per my knowledge, text detection models should work (given that supported operations are used), but had doubts about text recognition models as those are sequential models (RNN, LSTM, etc.). 

Can someone please clarify about this?

Thanks!

Sourabh

  • 0
    Prodigy 146 points

    Hi team,

    Can anyone please shed some light on this issue?

    Thanks and regards,
    Sourabh

  • 0 in reply to Sourabh Patil
    TI__Genius 11396 points

    As you know from the SDK documentation, we use onnxruntime as the inference runtime. So if you have an ONNX model, it can be run on our SDK and SoC. Certain layers are accelerated on DSP(C7x/MMA) using underlying TIDL, so when you select the model, you can select something with a backbone similar to the ones in our modelzoo (with typical layers such as Conv, BN, ReLU). The OCR detection head itself can be anything and will run on ARM using ONNX operators.

    For certain layers such as 2D Object Detection head, we have specialized layers in TIDL for maximum throughput. However, that is not strictly necessary. As I said above, if you have an ONNX model, it can be used.