TinyYoloV3 conversion from .pb to IRv10

[Tr__Ra]

Hello

Does the conversion from tensorflow frozen model of TinyYoloV3 to OpenVINO IR require the --reverse-input-channels flag for mo_tf.py? I know it is required for TF ssdv2 model. From my experiments, the Coco trained TinyYoloV3 seems to perform better when using the --reverse-input-channels flag.

Thank you!

[Sahira_Intel]

Hi Ra,

It is not required to use the --reverse_input_channels parameter when converting YOLOv3-Tiny. Below is the general conversion command:

python3 mo_t.f.py --input_model /path/to/yolo_v3_tiny.pb --tensorflow_use_custom_operations_config $MO_ROOT/extensions/front/tf/yolo_v3_tiny.json --batch 1

However, in general the color channel order (RG>B or BGR) of the input data should match the channel order of the model training dataset. 

For more information about when to reverse input channels, see this documentation.

I hope this information is helpful.

Sincerely,

Sahira 

[Tr__Ra]

To me it looks like the model performs better with reverse input channels.

My logic is that darknet, just like tf, uses rgb, while OpenVINO uses cv2 order of bgr