Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Kim__leila
Beginner
182 Views

Input size ERROR (when height and width are different)_yolov3

HI i have several questions about input size

i have trained darknet yolov3 with input size 192x96

1. when i dumped tensorflow model using tensorflow-yolo-v3, it made error because of input size.

so i changed anchors in yolo_v3.py and input size in convert_weight.py & convert_weight_pb.py

also i changed _upsample()function in yolo_v3.py like below 

else:
    new_height = out_shape[1]
    new_width = out_shape[2]
    #new_height = out_shape[2]
    #new_width = out_shape[1]

 

so i got ckpt file and pb file

 

2. and i run mo_tf.py with option --input_shape [1, 96,192,3]

so i got bin,mapping, xml files

 

3. Lastly, i am trying to inference my model using object_detection_demo_yolov3_async.py 

but it gives me a message that height and weight should be equal

Is there anyway i can use my own input size? not square shape input size 

0 Kudos
1 Reply
JesusE_Intel
Moderator
182 Views

Hi Leila,

Thanks for reaching out, the demo only supports square inputs. I have not tried using a non-squared input with YOLOv3 and the OpenVINO toolkit. You may have to modify the demo to meet your needs. However, I am uncertain if this will work. Is there a particular reason why you need to keep the input shape to 192 x 96?

Regards,

Jesus

 

Reply