Hei and sorry if I perhaps double-post this;
we have been looking for a fairly good OCR model, and we found one in easyocr.
We have been pleased to see that paddbleOCR has been included in the Zoo, and have been testing with it eagerly, however, it appears the mode has input range up to 960.
We tendentially work with 720p (1280x720p); what would be the best procedure to re-generate the paddleOCR with that given input resolution?
Thanks!
Hey @Tone_Kastlunger,
Welcome to the Hailo Community!
You’ve got a couple of options here. You can grab the ONNX model from the model zoo and recompile it with a resize layer added before the model so your input comes in at 720p. That way it’s all handled on the model side and on Hailo-8.
Alternatively, if you prefer to handle it separately, you can do the resizing in OpenCV before you feed the data into the OCR. Either approach works depending on your workflow!
Hope this helps!
Hi Omria, thank you for your kind reply. I saw posts with scripts for re-generating the model from scratch; how big should the dataset be?
It really depends on the model, but if you’re using PaddleOCR, I’d probably go with something between 500 and 1000. It’ll depend on what CPU and GPU you’re using for quantization and how much performance you need.