BBAI-Dried Sea Cucumber Species Identifier

RE: BBAI-Dried Sea Cucumber Species Identifier

mayermakes — Mon, 09 Dec 2019 20:54:07 GMT

words can express how muc hI love this project.

first for its specificness and second for its execution.

and on third I might add , I'ma huge bug nerd, so this use of AI is just sooo awesome!!!

RE: BBAI-Dried Sea Cucumber Species Identifier

shabaz — Thu, 05 Dec 2019 17:57:57 GMT

Very interesting, this was an impressive project.

Could you explain Darknet a little more if you don't mind? A video walk-through of that or your project would be awesome.

Anyway, excellent work.

RE: BBAI-Dried Sea Cucumber Species Identifier

three-phase — Sun, 24 Nov 2019 21:11:40 GMT

Quite an ambitious project that has produced some really great results. Well done.

Kind regards.

RE: BBAI-Dried Sea Cucumber Species Identifier

dubbie — Sun, 24 Nov 2019 18:08:27 GMT

The accuracy you have achieved with your neural network is pretty good. Obviously it would be nice if it was better. More training data should help. Getting good training data is sometimes quite difficult. I have tried recognising letters with an ANN, where there is really only one correct image (once the image has been normalised). Any other data is wrong so there is not much point on training the ANN on slightly wrong letters.

One thing I did which helped with the level of confidence of the selection made by the ANN was to get a percentage reliability value (I could not see whether your system produced this) as well as the second choice. I cannot remember the exact maths I used, but it was something to do with statistics - I'll see if I can find it somewhere, sometime. It was useful to see the difference between the first and second choices. A small difference would indicate an increased level of uncertainty in the accuracy of the first choice, so a first choice of 80% certain with a second choice of 20% would probably be more reliable than a 90% first choice and 80% second choice.

Dubbie

RE: BBAI-Dried Sea Cucumber Species Identifier

Sean_Miller — Sun, 24 Nov 2019 13:54:32 GMT

Great stuff! I wonder how one could take your trained model into the BBAI classification.tidl.cpp example code? That example makes use of the BBAI hardware.

Here is the section of the classification.tidl.cpp code that sets up the filters and config:

[embed:dc8ab71f-3b98-42d9-b0f6-e21e02a0f8e2:3ef06e6b-9b7d-4651-b12d-db8be7f5e991:type=c_cpp&text=++++std%3A%3Acout+%3C%3C+%22Initializing+filter%22+%3C%3C+std%3A%3Aendl%3B%0A++++populate_labels%28%22%2Fusr%2Fshare%2Fti%2Fexamples%2Ftidl%2Fclassification%2Fimagenet.txt%22%29%3B%0A++++selected_items_size+%3D+9%3B%0A++++selected_items+%3D+%28int+*%29malloc%28selected_items_size*sizeof%28int%29%29%3B%0A++++if+%28%21selected_items%29+%7B%0A++++++++std%3A%3Acout+%3C%3C+%22selected_items+malloc+failed%22+%3C%3C+std%3A%3Aendl%3B%0A++++++++return+false%3B%0A++++%7D%0A++++selected_items%5B0%5D+%3D+609%3B+%2F*+jeep+*%2F%0A++++selected_items%5B1%5D+%3D+627%3B+%2F*+limousine+*%2F%0A++++selected_items%5B2%5D+%3D+654%3B+%2F*+minibus+*%2F%0A++++selected_items%5B3%5D+%3D+656%3B+%2F*+minivan+*%2F%0A++++selected_items%5B4%5D+%3D+703%3B+%2F*+park_bench+*%2F%0A++++selected_items%5B5%5D+%3D+705%3B+%2F*+passenger_car+*%2F%0A++++selected_items%5B6%5D+%3D+779%3B+%2F*+school_bus+*%2F%0A++++selected_items%5B7%5D+%3D+829%3B+%2F*+streetcar+*%2F%0A++++selected_items%5B8%5D+%3D+176%3B+%2F*+Saluki+*%2F%0A++++std%3A%3Acout+%3C%3C+%22loading+configuration%22+%3C%3C+std%3A%3Aendl%3B%0A++++configuration.numFrames+%3D+0%3B%0A++++configuration.inData+%3D+%0A++++++++%22%2Fusr%2Fshare%2Fti%2Fexamples%2Ftidl%2Ftest%2Ftestvecs%2Finput%2Fpreproc_0_224x224.y%22%3B%0A++++configuration.outData+%3D+%0A++++++++%22%2Fusr%2Fshare%2Fti%2Fexamples%2Ftidl%2Fclassification%2Fstats_tool_out.bin%22%3B%0A++++configuration.netBinFile+%3D+%0A++++++++%22%2Fusr%2Fshare%2Fti%2Fexamples%2Ftidl%2Ftest%2Ftestvecs%2Fconfig%2Ftidl_models%2Ftidl_net_imagenet_jacintonet11v2.bin%22%3B%0A++++configuration.paramsBinFile+%3D+%0A++++++++%22%2Fusr%2Fshare%2Fti%2Fexamples%2Ftidl%2Ftest%2Ftestvecs%2Fconfig%2Ftidl_models%2Ftidl_param_imagenet_jacintonet11v2.bin%22%3B%0A++++configuration.preProcType+%3D+0%3B%0A++++configuration.inWidth+%3D+224%3B%0A++++configuration.inHeight+%3D+224%3B%0A++++configuration.inNumChannels+%3D+3%3B%0A++++configuration.layerIndex2LayerGroupId+%3D+%7B+%7B12%2C+2%7D%2C+%7B13%2C+2%7D%2C+%7B14%2C+2%7D+%7D%3B%0A++++configuration.enableApiTrace+%3D+false%3B%0A++++configuration.runFullNet+%3D+true%3B]

TI has a conversion tool to make the bin files that describe the model networks from TensorFlow. Unfortunately, their referenced website has been down for days that talks about it:

http://software-dl.ti.com/processor-sdk-linux/esd/docs/latest/linux/Foundational_Components_TIDL.html

See ya',

Sean

RE: BBAI-Dried Sea Cucumber Species Identifier

14rhb — Sun, 24 Nov 2019 12:43:21 GMT

I was wondering how you were getting on this your project [mention:c59ebc0edfa949458102394df67187e8:e9ed411860ed4f2ba0265705b8793d05] as things went a bit quiet, and now I can see why. This is great and a very interesting write up to follow through later in my own slower time.

If your accuracies of 65-85% were from only 20 images of each species then that is fairly impressive, well done. I guess if you can get a few of each dried species you could photograph them yourself at different rotations and lighting conditions to expand your dataset.