IR YOLOV3 don't recognize well

[Morishige__Akihito]

I trained original data on tensorflow YOLOV3, and convert to IR.

But it doesn't detect well comparing to original tensorflow YOLO.

Where should I change when modifying the number of class ?

(COCO dataset original: 80→change to 2class)

■My study

・YOLO model：https://github.com/mystic123/tensorflow-yolo-v3

・Training data: Original

・Convert IR：Refer https://software.intel.com/en-us/articles/OpenVINO-Using-TensorFlow

・Change：　Coco.names: 80class → 2class (XXX, XXX)

　　　　　　　Yolo_v3.json: classes 80→classes 2

    →I think I should change other parameter or model architecture, so could you tell me where I should change ?

[Shubha_R_Intel]

Dear Akihito, I hope this thread is of value to you: (the original poster decided in the end that it "..was a mistake in the logic of preprocessing and postprocessing".)

https://github.com/PINTO0309/OpenVINO-YoloV3/issues/2

[Shubha_R_Intel]

Also check this recent post. It seems pre-processing and post-processing is key.

https://software.intel.com/en-us/forums/computer-vision/topic/805531