Why yolov3 ir so big and performance is so pool ?

[nd1]

recently, I successfully convert the yolov3 to IR , but the new IR model size is too big.  my openvino version is computer_vision_sdk_2018.4.420.  I follow the doc directory :docs/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html  to convert.

But the step I follow seems to some error, I failed.  Then I modify the step as follow:

python mo_tf.py

--input_model yolo_v3.pb 

--tensorflow_use_custom_operations_config yolo_v3.json

--input_shape=[1,416,416,3]

--data_type=FP32

then the convertion success.

but I can not understand why that the models demo offered is so small and excellent performance,however the model I convert (yolov3)  performance is pool and the size is large .

Can anyone help me ?

[nikos1]

Hello Wang Nd,

 > performance is pool and the size is large .

YOLO v3 is deep so performance may be poor but will be more accurate.

If you need better inference speed you may want to try tiny YOLO.

Nikos

[nd1]

Hello Nikos,

Thanks for your reply.  I am also very interested in the small IR model the demos offered. eg:  vehicle-attributes-recognition, I  also test it. The performance is very Gei Li (Chinese words mean very excellent). I wander if I can also train the IR model directally (no convertion).  Because the document has no mention about the IR model coming from.  Currentlly I convert tensorflow and yolov3, I wish to know how to train such a small size model, so I can apply it to my application.

Wang Nd