Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Cookpadで何つくっている?

 Cookpadで何つくっている?

Leszek Rybicki

October 29, 2017
Tweet

More Decks by Leszek Rybicki

Other Decks in Research

Transcript

  1. ࣗݾ঺հ • 1990 BASIC, ASM, C • 1995 Pascal, VB

    • 2000 C++ • 2005 D • 2010 C, Python • 2012 PHP • 2015 JS • 2016 Python
  2. DCNN global pooling fully connected RGB x 240 x 240

    pixels 2048 features x 8 x 8 one-hot class vector 2048 features food plant person pet other
  3. DCNN global pooling fully connected RGB x 240 x 240

    pixels 2048 features x 8 x 8 one-hot class vector 2048 features DCNN global pooling 1x1 convo layer RGB x 240 x 240 pixels 2048 features x 8 x 8 one-hot class vector food not food not
  4. w %$//*ODFQUJPO7 w USBJOFEPOBMMGPPEBOE BMMOPOGPPEQIPUPT
 USBJOJOHJTSFBMMZGBTU  w BWFSBHFQPPMJOH w

    OPHMPCBMQPPMJOHMBZFS BGUFSUSBJOJOH w VTFSFTVMUJOHIFBUNBQ GPSDMBTTJpDBUJPO DCNN global pooling 1x1 convo layer RGB x 512 x 512 pixels 2048 features x 14 x 14 one-hot class vector food not
  5. ·ͱΊ •working on food images is an interesting challenge! •

    food / non-food: • multi-class classifier is better than two class • fully-convolutional classifier is better than multi-class • multi-class object detection from single-object training data