Guide for OpenVINO and DLStreamer deploy

[hungtrieu07]

Hello everyone. I'm looking for a guide for convert the PPOCRv3 text recognition model to OpenVINO format for my DLStreamer pipeline. My dataset just have Latin character and number from 0 to 9. Thanks for your help 😄

[GreatV]

To deploy the PPOCRv3 text recognition model using OpenVINO and DLStreamer, you need to follow these steps:

1. Convert PPOCRv3 Model to OpenVINO Format

Since OpenVINO works with models in IR format, you need to convert the PaddleOCR model using OpenVINO's model optimizer.

Steps to Convert PPOCRv3 to OpenVINO:

1. Export the PaddleOCR Model:

   • If you haven't already, export the trained PPOCRv3 recognition model to an inference model:

     python3 tools/export_model.py -c configs/rec/rec_mv3_none_bilstm_ctc.yml \
       -o Global.pretrained_model=your_model_path/best_accuracy Global.save_inference_dir=./inference/ppocrv3_rec

2. Convert PaddlePaddle Model to ONNX:
   Use paddle2onnx to convert the exported model:

   paddle2onnx --model_dir inference/ppocrv3_rec \
     --model_filename inference.pdmodel \
     --params_filename inference.pdiparams \
     --save_file ppocrv3_rec.onnx \
     --opset_version 11

3. Convert ONNX to OpenVINO IR Format:
   Use OpenVINO’s Model Optimizer:

   mo --input_model ppocrv3_rec.onnx --output_dir openvino_model

   This generates ppocrv3_rec.xml and ppocrv3_rec.bin.

---

2. Integrate PPOCRv3 Model in DLStreamer

DLStreamer is a GStreamer-based framework compatible with OpenVINO.

Steps to Deploy Recognizer in DLStreamer Pipeline:

1. Install OpenVINO & DLStreamer:

   sudo apt update && sudo apt install -y intel-openvino-runtime-ubuntu

2. Define a GStreamer Pipeline:
   You need gvadetect, gvaclassify, and other DLStreamer elements for the pipeline. An example pipeline:

   gst-launch-1.0 filesrc location=your_video.mp4 ! decodebin ! videoconvert ! \
     gvadetect model=openvino_model/ppocrv3_rec.xml ! \
     gvaclassify model=openvino_model/ppocrv3_rec.xml ! \
     videoconvert ! autovideosink

   Adjust model paths as necessary.

---

3. Model Optimization for Latin Characters and Numbers

Since your dataset consists of Latin characters and numbers (0-9), modify dict.txt and re-export the model before conversion. Ensure only the necessary characters are included to minimize model size.

---

Conclusion

1. Export PPOCRv3 model from PaddleOCR.
2. Convert to ONNX and then OpenVINO IR format.
3. Integrate with DLStreamer using GStreamer pipelines.

If you need further details on fine-tuning PPOCRv3 before conversion, you can refer to this discussion:
🔗 Advice on Optimal Configuration for Fine-Tuning PP-OCRv3

Let me know if you need additional clarification! 🚀

---

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.

[lemonschwein]

Hi GreatV, thanks for your answer here. I am facing the same issue, but with new problem, since the newest official PP-OCRv5 inference model only doesn't have .pdmodel file anymore. They replaced the .pdmodel files with IR format: .json, .xml, pdiparams. Could you please give me any hints, if you know the answer for my issue？ Thanks in advance!