Profile cover photo
Profile photo
Javier Barandiaran Martirena
About
Posts

Post has shared content
Add a comment...

Post has attachment
Casual 3D Photography
Peter Hedman, Suhib Alsisan, Richard Szeliski and Johannes Kopf
http://visual.cs.ucl.ac.uk/pubs/casual3d
https://youtu.be/wGBistgOsyQ
Add a comment...

Post has shared content
In this new work from the Dyson Robotics Lab at Imperial College, by +Jan Czarnowski, +Stefan Leutenegger and myself, we show that the image pyramid normally used for tracking (Lucas Kanade alignment) in dense SLAM systems such as DTAM can be simply replaced by the output of convolutional layers of a standard pre-trained CNN. The result is tracking which is much more robust to scene changes such as lighting. We show the results in a keyframe based pure rotation dense SLAM system inspired by +Steven Lovegrove Lovegrove's PhD work.

This is a neat result in itself, but is interesting in the longer term as we try to find dense but efficient ways of representing the appearance and shape of scenes for the next generation of SLAM systems. We believe that there are many useful levels of representation between raw pixel values and the human-defined object level maps which many semantic SLAM approaches are trying to build. The outputs of successive layers of a CNN trained for classification represent increasingly semantic entities with a built-in multi-scale character so give a good insight into this. Future SLAM systems will build generative maps which are able to predict the views of cameras which fly through them, but will not necessarily need to do this down to the photometric level, depending on what the maps are actually to be used for.

Still, the next step in moving to a completely learned scene representation which exists in 3D and can be used to render arbitrary generative views is very challenging, and will require us to move away from the comfortable rectangular feature arrays of standard CNNs.

This work will be presented in the ICCV 2017 workshop Geometry Meets Deep Learning. Paper available here:
https://arxiv.org/abs/1708.08844
Add a comment...

Post has attachment
NVIDIA DRIVENet Demo - Visualizing a Self-Driving Future
https://youtu.be/HJ58dbd5g8g
Add a comment...

Post has attachment

Post has attachment
Impressive work and demonstration Real-Time High-Fidelity Facial Performance Capture https://youtu.be/MMa2oT1wMIs #SIGGRAPH2015 paper available http://www.disneyresearch.com/publication/realtimeperformancecapture/  
Add a comment...

Post has attachment
“The greatest glory in living lies not in never falling, but in rising every time we fall.” Nelson Mandela
https://youtu.be/g0TaYhjpOfo
Add a comment...

Post has attachment
New advanced driver assistance system demo by Vicomtech-IK4
Add a comment...

Post has attachment

Post has attachment
Add a comment...
Wait while more posts are being loaded