Fine Tune BI6.0 AI OCR

[adventuredphoto]

I noticed that BI6 uses the global_mobile_vit_v2_ocr.onnx...... for LPR OCR. Is there a way to fine tune that using the previous model?
global_mobile_vit_v2_ocr Model Config · Issue #75 · ankandrew/fast-plate-ocr Seems like they don't support the old format and only the cct ones. Any guide or help would be appreciated

 

[MikeLud1]

I noticed that BI6 uses the global_mobile_vit_v2_ocr.onnx...... for LPR OCR. Is there a way to fine tune that using the previous model?
global_mobile_vit_v2_ocr Model Config · Issue #75 · ankandrew/fast-plate-ocr Seems like they don't support the old format and only the cct ones. Any guide or help would be appreciated

It would be better if Ken switches to using the cct-s-v1-global-model and/or cct-xs-v1-global-model, it has better accuracy. I have not had a chance to email Ken asking if he will make the switch.

Average plate accuracy (all 66 countries)​

Region‑level averages (plate accuracy)​

I grouped countries into 3 rough regions

• EU‑like: most European countries in your data (44 countries).
• Americas: Argentina, Brazil, Canada, Mexico, USA (5).
• Asia‑Pac: Armenia, Azerbaijan, Bahrain, Cambodia, Georgia, Hong Kong, Indonesia, Israel, Malaysia, Qatar, Singapore, Thailand, Turkey, Vietnam, Australia, New Zealand (16).

​

Summary​

• CCT‑S v1dominates:
  • Best average plate accuracy in all regions.
  • Wins in ~91% of countries on plate accuracy.
  • CCT‑XS v1 is consistently strong and occasionally ties or exceeds CCT‑S in a few countries.
• MobileViT v2 OCR generally underperforms both CCT variants, with a few local wins (Iceland, Slovakia).

 

[adventuredphoto]

Yeah that would be amazing… wonder if he can do it