jueves, 11 de diciembre de 2014

Very Deep CNN (19conv) first convolution filters 3D visualization

These are the filters of the first convolution layer of the very deep Convolutional Neural Network from Karen Simonyan and Andrew Zisserman available in the webpage www.robots.ox.ac.uk/~vgg/research/very_deep/ The authors have available an arXiv version of a paper in http://arxiv.org/abs/1409.1556

They got the 1st position for the localization task and the 2nd position in the classification task in ImageNet Challenge 2014. In order to get these results they evaluated different architectures with an increasing depth. This is the projection of the first convolutional layer filters in the RGB colorspace. Click on the images to see a 3D representation of the filters components.

As in the Alexnet example we can do a linear transformation of the original RGB channels and visualize the same colors in the YUV colorspace. In this case the distribution of the points is not that clean and it seems that the distribution of the colours is more spread. This could be because the number of weights is very reduced 64x3x3x3 + 64 = 1792, compared to Alexnet 96x3x11x11 +96 = 34944.

If you find these interesting you can take a look at the results in my Master Thesis: webpage or the pdf and do not hesitate to ask me any question.

No hay comentarios:

Publicar un comentario