c++ sdk running MMOCR tensorrt model using 500ms

[leemayi]

I trained the mmocr model using dbnet&satrn (small), and export tensorrt model

C++time statistics:
dbnet model : 30ms
satrn model: 480ms

export config:
mmdeploy-1.1.0\configs\mmocr\text-recognition\text-recognition_tensorrt_dynamic-32x32-32x640.py

ENV:
mmdeploy: main
mmocr: 1.0.0
mmdet: 3.0.0
tensort: 8.6.1.6

QA:
Can I optimize the textrecog model time to within 100ms?

[irexyc]

You shoud first analyze the time using of each part. You can refer to this doc

If the model inference time (trt inference without pre/post processing) already takes much than 100ms, I think use a smaller model could be a better choice