yolo_v3 died in 2019.R3

[ininder__gg]

I remove last version (openvino_2019.2.275) and re-installed current version (openvino_2019.3.334) , I find that my custom 11 classes tiny-yolo model was broken when using the omz_demos_build/object_detection_demo_yolov3_async, my environment is ubuntu 16.04, using NCS 1 as MYRIAD and the model is FP16.

To double check the problem, I use darknet weights and cfg and redo the document, My yolo_v3_tiny.json is

[
  {
    "id": "TFYOLOV3",
    "match_kind": "general",
    "custom_attributes": {
      "classes": 80,
      "anchors": [10, 14, 23, 27, 37, 58, 81, 82, 135, 169, 344, 319],
      "coords": 4,
      "num": 6,
      "masks": [[3, 4, 5], [0, 1, 2]],
      "entry_points": ["detector/yolo-v3-tiny/Reshape", "detector/yolo-v3-tiny/Reshape_4"]
    }
  }
]

and I use object_detection_demo_yolov3_async as:

./object_detection_demo_yolov3_async -m ~/test/frozen_darknet_yolov3_model.xml -i cam -d MYRIAD
 

it shows

InferenceEngine: 
    API version ............ 2.1
    Build .................. custom_releases/2019/R3_cb6cad9663aea3d282e0e8b3e0bf359df665d5d0
    Description ....... API
[ INFO ] Parsing input parameters
[ INFO ] Reading input
[ INFO ] Loading Inference Engine
[ INFO ] Device info: 
    MYRIAD
    myriadPlugin version ......... 2.1
    Build ........... 30677
[ INFO ] Loading network files
[ INFO ] Batch size is forced to  1.
[ INFO ] Checking that the inputs are as the demo expects
[ INFO ] Checking that the outputs are as the demo expects
[ INFO ] Loading model to the device
[ INFO ] Start inference 
 

and the result gose strange

https://imgur.com/a/aid6Fib

Not sure what's going on but the last version works fine.

[Shubha_R_Intel]

Dearest ininder, gg,

Indeed that picture looks messed up. "Died" is an understatement ! This is very important though. Did you regenerate your IR on R3 ? You cannot use the previous IR. I guess you must have re-generated the IR because you said above that you redid the document. Did you also use the R3 versions of the *.json ? Also you say you used NCS1. Would you happen to have an NCS2 which you could try ?

Please let me know,

Shubha

[ininder__gg]

hi sir,

    actually I use virtual box to build a new ubuntu computer to test the R3, so the IR model should be built in R3 suitable environment since there is no other vino. I have no NCS 2 to test. Dose NCS 1 not be supported in openvino?

    I also take the *.pb file back to last version (R2) and use R2's converter to test last version's object_detection_demo_yolov3_async, it works very well

here's the result:

R2:

FP16 in NCS 1

FP32 in CPU

R3:

FP16 in NCS 1

FP32 in CPU

In R3, the yolov3-tiny works slightly different from the R2 in CPU, but the result is somehow reasonable, however, the result in FP16 in NCS1 is totally hard to understand. 

p.s. By the way, in R3 the object_detection_demo_yolov3_async demo can also run FP16 model in CPU and runs FP32 model in MYRIAD without any warning which is very strange. MYRIAD should not be allowed to run the FP32 model.

[Shubha_R_Intel]

Dear ininder, gg

Thank you for all the details. I will try to reproduce this myself on R3. And yes, NCS1 is supported by OpenVino today.

Thanks for your patience !

Shubha

[Shubha_R_Intel]

Dear ininder, gg,

You are correct. Yolo V3 does not work properly in 2019R3 for NCS1. I reproduced this myself. It works fine on NCS2 though. I will file a bug on your behalf.

Sorry for the inconvenience !

Shubha

[Søeberg__Philip]

Shubha R. (Intel) wrote:

I will file a bug on your behalf.

Is the bug tracking system internal or can we follow its progress? I am facing same issue with NCS1. I bought a NCS2 and it work just fine with 2019.R3 but swapping NCS2 with NCS1 corrupts the output.

[Shubha_R_Intel]

Dear Søeberg, Philip

 Yes unfortunately the bug tracking is entirely internal. 

Sorry about that.

Shubha