Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Tsin__Ross
New Contributor I
78 Views

Don't get disired output on OpenVINO/NCS2 with same tiny yolov3 model.

When run

darknet.exe  detector test areca-c.data yolov3-tiny.cfg backup/yolov3-tiny_13500.weights E:\AI\Data\areca-train\train\DJI_0775_2.jpg

I got these:

================================================

layer     filters    size              input                output
   0 conv     16  3 x 3 / 1   416 x 416 x   3   ->   416 x 416 x  16 0.150 BF
   1 max          2 x 2 / 2   416 x 416 x  16   ->   208 x 208 x  16 0.003 BF
   2 conv     32  3 x 3 / 1   208 x 208 x  16   ->   208 x 208 x  32 0.399 BF
   3 max          2 x 2 / 2   208 x 208 x  32   ->   104 x 104 x  32 0.001 BF
   4 conv     64  3 x 3 / 1   104 x 104 x  32   ->   104 x 104 x  64 0.399 BF
   5 max          2 x 2 / 2   104 x 104 x  64   ->    52 x  52 x  64 0.001 BF
   6 conv    128  3 x 3 / 1    52 x  52 x  64   ->    52 x  52 x 128 0.399 BF
   7 max          2 x 2 / 2    52 x  52 x 128   ->    26 x  26 x 128 0.000 BF
   8 conv    256  3 x 3 / 1    26 x  26 x 128   ->    26 x  26 x 256 0.399 BF
   9 max          2 x 2 / 2    26 x  26 x 256   ->    13 x  13 x 256 0.000 BF
  10 conv    512  3 x 3 / 1    13 x  13 x 256   ->    13 x  13 x 512 0.399 BF
  11 max          2 x 2 / 1    13 x  13 x 512   ->    13 x  13 x 512 0.000 BF
  12 conv   1024  3 x 3 / 1    13 x  13 x 512   ->    13 x  13 x1024 1.595 BF
  13 conv    256  1 x 1 / 1    13 x  13 x1024   ->    13 x  13 x 256 0.089 BF
  14 conv    512  3 x 3 / 1    13 x  13 x 256   ->    13 x  13 x 512 0.399 BF
  15 conv     18  1 x 1 / 1    13 x  13 x 512   ->    13 x  13 x  18 0.003 BF
  16 yolo
  17 route  13
  18 conv    128  1 x 1 / 1    13 x  13 x 256   ->    13 x  13 x 128 0.011 BF
  19 upsample            2x    13 x  13 x 128   ->    26 x  26 x 128
  20 route  19 8
  21 conv    256  3 x 3 / 1    26 x  26 x 384   ->    26 x  26 x 256 1.196 BF
  22 conv     18  1 x 1 / 1    26 x  26 x 256   ->    26 x  26 x  18 0.006 BF
  23 yolo
Total BFLOPS 5.448
Loading weights from backup/yolov3-tiny_13500.weights...
 seen 64
Done!
E:\AI\Data\areca-train\train\DJI_0775_2.jpg: Predicted in 10.122000 milli-seconds.
areca: 100%
areca: 100%
areca: 100%
areca: 100%
areca: 100%
areca: 100%
areca: 100%
areca: 100%
areca: 100%
areca: 100%
areca: 100%
saving to DJI_0775_2.jpg

DJI_0775_2.jpg

 

However, with OpenVINO/NCS2, I got this

output.jpg

Is this because we have different layers in IE model compared to darknet model?

0 Kudos
3 Replies
Shubha_R_Intel
Employee
78 Views

Dear Ross,

Please use PINTO's github repo for inspiration :

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

My guess is that your anchor settings are incorrect.

Thanks for using OpenVino !

Shubha

Tsin__Ross
New Contributor I
78 Views

Shubha R. (Intel) wrote:

Dear Ross,

Please use PINTO's github repo for inspiration :

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

My guess is that your anchor settings are incorrect.

Thanks for using OpenVino !

Shubha

 

Hi Shubha,

I tried, but seemed not working,

check it out

pinto.jpg

Tsin__Ross
New Contributor I
78 Views

I am going to give up openvino....

too much time focusing on model integration...

Reply