Intro to computer vision

Иван Талалаев IT meetup Dev_stav 22.09.2018

Transcript

1. Stv 2018 Intro to Computer Vision
2. About me
3. Вкусвилл
4. Uniform control Ложных 3,02% Верно 85,96%
5. Empty shelves Ложных 6,1% Верно 94,91%
6. Supply time Ложных 2,2% Верно 86,62%
7. Broken monitors Ложных 0,5% Верно 90,27%
8. Queues Статистика по очередям ведется в БД 9:00 22:00 Найденные покупатели выделяются боксами на исходном изображении Кассиры и покупатели имеют различные классы, что позволяет не учитывать их в статистике по покупателям Ложных 7,2% Верно 85,96%
9. Some history • 1958 Perceptron by Frank Rosenblatt Y = xA + b
10. Some history
11. Some history • 1969 M. Minsky – perceptron restrictions Just linear classifier – even cannot resolve XOR problem
12. Some history
13. Some history • 1969 M. Minsky – perceptron restrictions
14. Some history • More deeper…
15. Some history • 1974 J. Verbos / A.I. Galushkin – backward propagation • 1982 Hopfield network, T. Kohonen self-organizing map • 1986 G. Hinton / S. I. Bartcev – backward propagation • 2007 G. Hinton Restricted Boltzmann Machine •
16. Convolution Neural Networks • 2012 AlexNet (Alex Krizhevsky, Geoffrey Hinton, and Ilya Sutskever) • ImageNet Large Scale Visual Recognition Challenge • The network achieved a top-5 error of 15.3%, more than 10.8 percentage points ahead of the runner up • • Top – 1: 62.5% Top – 5: 84.7%
17. Convolution Neural Networks • Convolutional neural network
18. Convolution Neural Networks • AlexNet 60 million parameters 650 000 neurons 5 convolutional layers 3 fully connected layers
19. Convolution Neural Networks
20. Convolution Neural Networks
21. Convolution Neural Networks • Pooling after convolution
22. Convolution Neural Networks
23. Convolution Neural Networks • Fully connected layers on top of convolution
24. Convolution Neural Networks • Fully connected layers
25. Convolution Neural Networks • Cross entropy loss
26. Convolution Neural Networks • Cross Entropy Loss
27. Convolution Neural Networks • Error Back Propagation with SGD E = 0.86
28. Convolution Neural Networks • Gradient Descent •
29. Convolution Neural Networks • Gradient Descent •
30. Convolution Neural Networks
31. Convolution Neural Networks
32. Convolution Neural Networks
33. Convolution Neural Networks