[HN Gopher] Total Relighting: Learning to Relight Portraits for ...
       ___________________________________________________________________
        
       Total Relighting: Learning to Relight Portraits for Background
       Replacement
        
       Author : Hard_Space
       Score  : 38 points
       Date   : 2021-05-09 07:04 UTC (1 days ago)
        
 (HTM) web link (augmentedperception.github.io)
 (TXT) w3m dump (augmentedperception.github.io)
        
       | nakedgremlin wrote:
       | Incredible work. Coming from a POV of a person who has to wrangle
       | a lot of VFX heavy creative, it's great to see the technical
       | walkthroughs like this, especially the workflow overviews to see
       | how it matches with our render process.
        
       | tqi wrote:
       | Interesting - as I understand it, one of the big tech innovations
       | from Disney's The Mandalorian was LED projected virtual sets that
       | solved the lighting issues from green screens (see ~2:10 mark of
       | this video: https://www.youtube.com/watch?v=Ufp8weYYDE8), making
       | VFX scenes cheaper and faster to produce in exchange for higher
       | up front costs. It had the added benefit of giving actors more to
       | work off of.
       | 
       | I wonder if this technique will swing the cost pendulum back in
       | favor of traditional green screens?
        
       | dheera wrote:
       | Is there code for this? There seems to be only a website in the
       | repo:
       | 
       | https://github.com/augmentedperception/total_relighting
        
         | ipsum2 wrote:
         | Don't expect code from this group, because their work usually
         | goes into Google Pixel camera software.
        
           | sorenjan wrote:
           | What is the legal situation for implementing things from
           | these kinds of papers? Am I allowed to use other's research
           | papers to make features in my own products, or should I
           | assume it's patented? I don't think copyright applies if I
           | write the code myself, right? Or are neural net models
           | copyrightable?
           | 
           | In this case you need specialized data from a lighting rig,
           | but what about in general when you can use your own or open
           | data to do the training, or in the cases when no machine
           | learning is used?
        
             | dheera wrote:
             | Also of note -- building that lighting rig is actually not
             | that hard if you have a spare walk-in closet and a bunch of
             | WS2812 lights. You can programmatically turn on and off
             | combinations of lights and automate the photography. But
             | you'll probably want several subjects of multiple races and
             | genders if you want it to work reliably though.
             | 
             | There's also this free dataset, although I haven't dug into
             | it to see how good it is:
             | 
             | https://drive.google.com/drive/folders/10luekF8vV5vo2GFYPRC
             | e...
             | 
             | ( and the associated paper:
             | https://zhhoper.github.io/dpr.html )
             | 
             | Are there any decent collections of human portrait Blender
             | scenes? If so lighting would be easy to change in a pinch,
             | and programmatically create a simulated dataset. A
             | collection of images of this quality would be awesome:
             | 
             | https://blenderartists.org/t/felicia-blend/701475
        
             | ipsum2 wrote:
             | It's probably patented (patent pending, because they take
             | forever to get approved), copyright doesn't apply if you
             | write the code and train the model yourself afaik
        
       ___________________________________________________________________
       (page generated 2021-05-10 23:01 UTC)