Description of a fast and stable Motion blur algorithm

Hello everyone, this article is a summary of another article: Guertin, Jean-Philippe & McGuire, M. & Nowrouzezahrai, D.. (2014). A fast and stable feature-aware motion blur filter. High-Performance Graphics 2014, HPG 2014 – Proceedings. 51-60.

It describes the algorithm for the Motion blur effect, which was proposed by the authors of the original article.

So, the algorithm itself:

The image is split into tiles for maximum blur radius of image space which ensures that each pixel is affected by no more than its surroundings of the tile.

*Pixel velocity vectors. Green shows the dominant speed in the tile.*

Each pixel has its own speed calculated:

$V[x,y]={v(p_{x,y})\over 2r} * {max(min)|v(p_{x,y})|*E,r),0.5)\over |v(p_{x,y })|+ε} +0.5$

Where $ε = 10^{-5}$ A — exposure time in seconds.

Each tile is assigned one dominant neighborhood speed $v_{max}$
Dominant velocity vectors in the tile. Green shows the dominant velocity vector among neighboring tiles
Pixel () and dominant ( $v_{max}$ ) pixel speeds p are processed as follows: a number of samples are placed along the central direction, interpolating between the normalized directions and perpendicular to the dominant $v_{max}^⊥$ as the pixel speed decreases to the minimum user threshold:
$v_c(p)=lerp(norm(v(p)),v_{max}^⊥,(||v(p)||-0.5)/γ$
Sampling directions $v_{max}$ (green), $v_{max}^⊥$ (yellow) and (blue)
On each tile, the angular difference between the maximum speeds of the tile and its neighbors is calculated: