Profile

Cover photo
Nick Sergievskiy
18,732 views
AboutPostsPhotosYouTube

Stream

Nick Sergievskiy

commented on a video on YouTube.
Shared publicly  - 
 
Magic!
1
Add a comment...

Nick Sergievskiy

commented on a video on YouTube.
Shared publicly  - 
 
Чел в метро действительно прикольный, пару раз видел, настроение поднимается от его позитивной речи
 ·  Translate
2
Add a comment...

Nick Sergievskiy

commented on a video on YouTube.
Shared publicly  - 
 
What about occlusion and deformation (rotate head 3 angles)?
1
Add a comment...
 
Intresng that best resault in Face Recognition is synchronously published Face++ (http://arxiv.org/abs/1403.2802) and facebook ( https://www.facebook.com/publications/546316888800776/ )
0.9727 ± 0.0065 vs 0.9725 ± 0.0081
14
6
Alberto Albiol's profile photoNegar Hassanpour's profile photoSoumith Chintala's profile photoMilton Wong's profile photo
 
Face++ used CNN with 6M parameters.
FB used NN with 120M parameters(no weight sharing).
Both showed similar accuracy result - speaking of power of convolutions...

But mine question, anyone had a chance to look at F++ paper? How Pyramid CNN is different from ordinarily CNN? F++ just trained CNN with one convolutional layer, then another CNN with two conv layers, but output of first c-layer shared in two layered CNN, then 3 layer CNN reuse output of first 2 convolutional layers from previous CNN... With  classifiers of each CNN trained independently, is that is what going on?
Add a comment...
 
Is there a deep learning methods that uses feature point (SURF, SIFT, ORB) with its scale factor and descriptor? OR why neural networks not interacts with standard computer vision algorithms?
1
2
अभिनेश्वर तोमर's profile photoMatt Kuenzel's profile photoLukas Mach's profile photoCristian Balint's profile photo
6 comments
 
If I see you question correctly, I do not see any restriction to feed raw features (extracted with CV methods) to any Deep Algorithm. In that way you are still possible to learn combinatorial information (correlations in other perspective) from those features. Also other than learning the features, you might use pre-training step to find good start point for your NN learning like in the case of Deep Belief Nets
Add a comment...

Nick Sergievskiy

Shared publicly  - 
 
Замечательная партия =)
 ·  Translate
1
Add a comment...
Basic Information
Gender
Male