Deep learning and neuron visualization

Currently teaching my network to go from a simple representation of a neuron morphology to a more detailed and realistic representation. Here are the first (promising) results:

Global illumination / Volume rendering applied to molecular visualization

Global illumination / Volume rendering

Last night, I got inspired by the Exposure Render: An Interactive Photo-Realistic Volume Rendering Framework, and started adding global illumination to volume rendering in Brayns. The first results are the following ones, and the code is about to be merged :-)

When Virtual Drums meet Interactive Brain Visualization

Could not help it, had to synchronize my VDrums with Brayns, the large scale interactive brain visualizer I created at EPFL.

Components used for the demo:
- Mido: http://mido.readthedocs.io/en/latest/ 
- Brayns: https://github.com/BlueBrain/Brayns

In the insideHPC News!

 In this silent video from the Blue Brain Project at SC16, 865 segments from a rodent brain are simulated with isosurfaces generated from Allen Brain Atlas image stacks. The work is derived from the INCITE program’s project entitled: Biophysical Principles of Functional Synaptic Plasticity in the Neocortex.

I produced 2 sequences of that video using Brayns, the application I designed in the context of the Blue Brain Project.

Happy New Year!!!

Happy new year, best wishes to all, and ray-tracing forever!

Brayns for neuro-robotics

Last tuesday, I presented how Brayns could be used to render high quality images for our colleagues from the neuro-robotics team. Brayns is hardware agnostic and it takes no more than one command line argument to switch between OSPRay (CPU) and OptiX (GPU) backends. The following video shows Brayns in action, on a 24MPixel display wall! Brayns is running on 1 machine powered by 2 Quadro K5000 NVIDIA GPUs.

SIMD accelerated voxelizer

SIMDVoxelizer is a CPU-based voxalizer taking advantage of vectorization units to speed up creation of 8bit raw volumes.

usage: SIMDVoxelizer <voxel_size> <cutoff_distance> <input_file> <output_file>

Input file is a binary array of floats: x, y, z, radius and value of elements. Each voxel of the final volume contains the sum of all elements with a weight that correspond to the value of the element divided by its squared distance to the voxel. Note that in the final volume, values are normalized.

This is currently a brute force implementation that produces accurate 8bit volumes.

The <output_file> is suffixed by the size of the volume.

SIMDVoxelizer makes use of the Intel ISPC compiler and requires ispc to be in the PATH.

To build SIMDVoxelizer, simply run make in the source folder.

Source code available on github.

Volume rendering and raytracing ... merged!

It's there, merged and operational in Brayns, and the first results are quite exciting.

Mixing volume rendering and ray-tracing allows more control on atmospheric effects which, in the case of neurosciences, can be used to represent the electromagnetic fields surrounding the neurons.

Volume Rendering coupled to Ray-Tracing: a win-win combination

I am currently working at adding Volume Rendering to the existing Brayns implementation. Using volumes clearly is the way to go to represent the activity that takes place outside of the geometry. Ray-tracing, on the other hand, concentrates on high quality surface rendering, processing shadows and other global illumination.

Many thanks to my colleagues  Raphael and Grigori without whom that development would have taken a few more days!


Currently in my volume branch, but soon to be merged into master!!