why after yolov3 trained model converted to IR, the binding box becames not accurate in location compared to yolov3.weights ir?
I have made a comparison. I first use other train data to train my own model with yolov3 neural networks, and then convert it to IR model.
the train classes include the person class. After that,I pick a picture from yolov3 official website named person.jpg , then
I use the IR model to check the person.jpg ,I find the binding box around the person not accurate—— The head,neck and under knee are not include
in the binding box. But when I use the yolov3.weights converted to IR to check the person.jpg ,the binding box around the person
is accurate.
the I use the meta pb(tensorflow) model to check person.jpg, the result is the same.
So I get a conclusion that the .weights convert to tensorflow pb model has problems. But the yolov3.weights converted to IR to check image ,the binding box is accurate, it's Puzzling ......
The image I post here : https://github.com/mystic123/tensorflow-yolo-v3/issues/78
the convert tool refers to here: https://github.com/mystic123/tensorflow-yolo-v3
can anyone help , much thanks !
