The inference result of mobilenet example on ncappzoo seems not correct

[idata]

I ran "python3 run.py" in the folder of mobilenet of ncappzoo, which infers a picture of "nps_electric_guitar.png", but the result is:

Mobilenet on NCS

653 military uniform 0.9873

765 rifle 0.0049477890 violin, fiddle 0.003397727 plane, carpenter's plane, woodworking plane 0.0014963414 assault rifle, assault gun 0.00078201

Finished

Is that correct?

[idata]

@xhuan28 I am observing the same thing. In fact the results are always the same no matter which image you point to. Will investigate further, maybe something is hardcoded there.

[idata]

Hi guys, has this problem been solved yet?

I've tried moblienet code provided by ncappzoo recently, and this situation also occurs.I've also tried doing inference on different pictures, the result are all wrong and weird.No matter what parameter I chose (version of mobilenet model, img_size, mean, std variable in run.py), the situation never changed.However, every other models (e.g. inception-v1, v2 , etc.) are all fine.

Does anyone have any suggestions or solutions? Thanks!

[idata]

@Thanatine @xhuan28 Thanks for bringing your concern to the forum. Try using the latest NCSDK (version 1.12.xx) while using the command: make run to run the Mobilenets example. Please let me know if this fixes your issue.

[idata]

@Tome_at_Intel I tried with 1.12.xx version, and the issue still exist.

For Eg. using the 512_Amplifier image, inference is incorrect.

Start download to NCS…

Mobilenet on NCS for image ../../data/images/512_Amplifier.jpg

808 solar dish, solar collector, solar furnace 0.92432

908 wine bottle 0.028793653 military uniform 0.0121900 water jug 0.012001412 apron 0.0052032

[idata]

@xhuan28, @xhuan28, @Thanatine, I tried these instructions with NCSDK 1.2.00 and master branch of ncappzoo. Let me know if this does not work for you.

NOTE: Steps 1 & 2 are essentially a breakdown of what happens when you run make run in ~/workspace/ncappzoo/tensorflow/mobilenets. Just make sure you have done a git pull inside your ~/workspace/ncappzoo/ directory.

1. Build mobilenet graph

cd ~/workspace/ncappzoo/tensorflow/mobilenets make

I get the following results

... Exporting GraphDef file... ... Freezing model for inference... ... Profiling the model... ... 29 MobilenetV1/Predictions/Softmax 0.0 19.1 0.199 --------------------------------------------------------------------------------------------- Total inference time 39.29 --------------------------------------------------------------------------------------------- Generating Profile Report 'output_report.html'... Movidius graph generated! You can run inferences using ncappzoo/apps/image-classifier project.

2. Run an inference using image-classifier app

cd ~/workspace/ncappzoo/apps/image-classifier python3 image-classifier.py --graph ../../tensorflow/mobilenets/model/graph --labels ../../tensorflow/mobilenets/model/labels.txt --mean 127.5 --scale 0.00789 --dim 224 224 --colormode="RGB" --image ../../data/images/512_Amplifier.jpg

I get the following results

============================================================== Top predictions for 512_Amplifier.jpg Execution time: 39.28766ms -------------------------------------------------------------- 94.8% 755:radio, wireless 3.9% 849:tape player 0.8% 486:CD player 0.3% 483:cassette player 0.0% 633:loudspeaker, speaker, speaker unit, loudspeaker system, speaker system ==============================================================

You should also see an visual representation of 512_Amplifier.jpg. I tried with a couple other images in data/images, they all worked fine.

[idata]

@AshwinVijayakumar , Yes I updated to the latest version of ncapp zoo and it is working now with NCSDK1.12. Thanks for your support.