Faster-rcnn only has shape is [1,3] ?

[AUNTP]

This my code >>>

from openvino.inference_engine import IENetwork
from openvino.inference_engine import IEPlugin
import numpy as np
import time
import cv2
import imutils

plugin = IEPlugin(device="CPU")

net = IENetwork(
model="handges.xml",
weights="handges.bin")

print("[INFO] preparing inputs...")
inputBlob = next(iter(net.inputs))
outputBlob = next(iter(net.outputs))

net.batch_size = 1

print(net.inputs[inputBlob].shape)
(n, c, h, w) = net.inputs[inputBlob].shape

#Out is 1,3 

but when I convert to IR the shape is [1,600,600,3]

and in xml file >> 

<?xml version="1.0" ?>
<net name="hand" version="10">
<layers>
<layer id="0" name="Preprocessor/mul/x/Output_0/Data_/copy_const" type="Const" version="opset1">
<data element_type="f32" offset="0" shape="1,1,1,1" size="4"/>
<output>
<port id="1" precision="FP32">
<dim>1</dim>
<dim>1</dim>
<dim>1</dim>
<dim>1</dim>
</port>
</output>
</layer>
<layer id="1" name="image_tensor" type="Parameter" version="opset1">
<data element_type="f32" shape="1,3,600,600"/>
<output>
<port id="0" precision="FP32">
<dim>1</dim>
<dim>3</dim>
<dim>600</dim>
<dim>600</dim>
</port>
</output>
</layer>

or I miss somethings 

Thankyou. 

[AUNTP]

And the last question when I was converting a model I didn't have use --input = image_tensor. Should I use  --input=image_tensor when I convert?

[Iffa_Intel]

Greetings,

Let's take a look at this faster rcnn model .prototxt file as an example since I don't have your specific .prototxt file (they should have the same concept): https://raw.githubusercontent.com/rbgirshick/py-faster-rcnn/master/models/pascal_voc/VGG16/faster_rcnn_end2end/test.prototxt

The first layer is an input layer and it's expected data is to be by size [1,3,255,255]. The input image size required in this case is 255x255. The image is an RBG image, hence 1 image require size 3 dimension . This is why the shape[1,3] is used.

You can do resizing to the model which you can refer to this step by step guide : https://www.youtube.com/watch?v=Ga8j0lgi-OQ

For simplicity, you can define the input_model as a variable pointing to your faster rcnn model.

If you want to use an input image with different size without training a new model,

the model optimizer can manipulate the sizes for you by using for example: --input_shape[1,3,100,100]

By giving this parameter, your image size will become 100x100. The sizes of all other layers down the pipes are also adjusted.

Sincerely,

Iffa

[Iffa_Intel]

Greetings,

Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question

Sincerely,

Iffa