I'm kind of surprised that Microsoft isn't using the acceleration and magnetic sensors in the phone to help determine the camera position. It's one of the features that phone cameras have that DSLR's don't.
Actually they do. Fig.2 in the paper, where the IMU output is used to refine the camera pose estimated by purely image based means.
Its true that the concept of 3D reconstruction from dense stereo/structure from motion is not new. However, the computational pipeline integrated entirely into a mobile device without any expensive hardware or offloading computation on a cloud, differentiates this effort from the previous ones.
Most deep learning algorithms used for image classification tasks use a data augmentation step - wherein they alter the training image through scaling, translation, etc.
According to the paper published here:http://arxiv.org/abs/1501.02876, they do additional transformations in the training images to make the learned model even more robust.
So the risk of using up cpu cycles on the same data again and again is reduced.
I have been using LInux since 2000. However, I always keep one machine dual boot. This is mainly due to work as my PhD advisor needs me to create presentations in M$ powerpoint which he can edit on his Win7. PPTs made on Openoffice Impress (I am not saying the native format) have issues embedding videos. Also some tables just don't work properly.
At home I have converted my wife to use Ubuntu.;-)
Slashdot Mirror
Well isn't Shannon's theory (more specifically the constraints on channel bandwidth) is already been replaced by Compressive Sensing?
A couple years ago - I stumbled on this. Seems very relevant to the discussion.
Writting "around 500,000" is quite controversial:
I don't think I wrote "around 500,000". AFAIR, I quoted the number from BBC.
I'm kind of surprised that Microsoft isn't using the acceleration and magnetic sensors in the phone to help determine the camera position. It's one of the features that phone cameras have that DSLR's don't.
Actually they do. Fig.2 in the paper, where the IMU output is used to refine the camera pose estimated by purely image based means.
Its true that the concept of 3D reconstruction from dense stereo/structure from motion is not new. However, the computational pipeline integrated entirely into a mobile device without any expensive hardware or offloading computation on a cloud, differentiates this effort from the previous ones.
Most deep learning algorithms used for image classification tasks use a data augmentation step - wherein they alter the training image through scaling, translation, etc. According to the paper published here:http://arxiv.org/abs/1501.02876, they do additional transformations in the training images to make the learned model even more robust. So the risk of using up cpu cycles on the same data again and again is reduced.
I have been using LInux since 2000. However, I always keep one machine dual boot. This is mainly due to work as my PhD advisor needs me to create presentations in M$ powerpoint which he can edit on his Win7. PPTs made on Openoffice Impress (I am not saying the native format) have issues embedding videos. Also some tables just don't work properly. At home I have converted my wife to use Ubuntu. ;-)