Yolov3 conversion to OpenVino IR accuracy loss

[Jkong]

According to this doc, OpenVino do have their own "Region" layer which supposedly to negate the accuracy loss during the conversion from Darknet to TF.

https://docs.openvino.ai/latest/openvino_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html

However, I am still looking at a significant accuracy loss. Is this still to be expected?

[Wan_Intel]

Hi Jkong,

Thanks for reaching out to us.

For your information, to convert YOLOv3 TensorFlow model into Intermediate Representation, you have to use yolo_v3.json or yolo_v3_tiny.json (depending on a model) configuration file with custom operations located in the <INSTALL_DIR>/tools/model_optimizer/extensions/front/tf repository.

Please refer to Convert YOLOv3 TensorFlow Model to IR for more information.

Regards,

Wan

[Jkong]

Hi,

As mentioned in the opening, the exact same document have been followed.

[Wan_Intel]

Hi Jkong,

May I know which demo applications are you using with the Intermediate Representation? Could you please share your model and step to convert your model into Intermediate Representation with us for further investigation?

Regards,

Wan

[Wan_Intel]

Hi Jkong,

Thank you for your question.

If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.

Regards,

Wan