• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Researchers Note Fundamental Ambiguity in Image Decomposition

    Replies examine limits of breaking color images into separate scene elements.

    AN
    KC
    2 Sources, 28d ago, first seen 28d ago

    TLDR

    Keenan Crane, a CMU researcher focused on geometry processing, replied that color images and video cannot be uniquely separated into geometry, materials, illumination, and camera parameters. He added that additional data will not resolve the issue. Alex Nichol, an OpenAI research engineer, replied that closer camera angles might improve geometry estimates from depth outputs of a model but said he cannot prove the approach works.

    Combined views

    3.7K

    2 Sources, first seen 28d ago

    Combined views

    3.7K

    2 Sources, first seen 28d ago

    33 likes
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    33 likes
    3 comments
    21 saves
    3 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    3 comments
    21 saves
    3 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    2 Sources

    @keenanisaliveExactly—the fundamental ambiguity (which @BenMildenhall and every other vision researcher is well aware of) is that color images/video are not in general uniquely decomposable into geometry, materials, illumination, and camera. No matter how much you believe in the bitter lesson, more data cannot overcome a problem that is ill-posed. The only thing that can do that is additional constraints or priors. (That is effectively why you get well-defined geometry with models like Meshy: the prior is built into the output representation.)
    @unixpickle@keenanisalive @mrbadrinath @fhahlbohm @chrisoffner3d @pzpzpzp1 @bmarkmiller @BenMildenhall I'm guessing you can get the model to give you better geometry by giving it camera angles closer to the surface you care about and looking at its depth outputs. Can't prove it tho

    2 Sources

    @keenanisaliveExactly—the fundamental ambiguity (which @BenMildenhall and every other vision researcher is well aware of) is that color images/video are not in general uniquely decomposable into geometry, materials, illumination, and camera. No matter how much you believe in the bitter lesson, more data cannot overcome a problem that is ill-posed. The only thing that can do that is additional constraints or priors. (That is effectively why you get well-defined geometry with models like Meshy: the prior is built into the output representation.)
    @unixpickle@keenanisalive @mrbadrinath @fhahlbohm @chrisoffner3d @pzpzpzp1 @bmarkmiller @BenMildenhall I'm guessing you can get the model to give you better geometry by giving it camera angles closer to the surface you care about and looking at its depth outputs. Can't prove it tho