@wjakob@fosstodon.org avatar

wjakob

@wjakob@fosstodon.org

Assistant professor leading EPFL's Realistic Graphics Lab. I work on inverse graphics, material appearance modeling and physically based rendering.

This profile is from a federated server and may be incomplete. Browse more on the original instance.

wjakob, to random
@wjakob@fosstodon.org avatar

I just released nanobind v2.0.0 featuring significantly improved typing infrastructure. It can now automatically build reference-quality stubs for your extensions (enabling completion in VS Code, static type checking via MyPy/PyLance, etc.) — https://nanobind.readthedocs.io/en/latest/changelog.html#version-2-0-0-may-23-2024

wjakob, to random
@wjakob@fosstodon.org avatar

Apple MLX switched from pybind11 to nanobind and got a nice performance improvement as a result. 🥳 https://github.com/ml-explore/mlx/pull/839

wjakob, to random
@wjakob@fosstodon.org avatar

If you try to optimize geometry using a differentiable renderer, there is an elephant in the room: geometry causes discontinuous visibility changes, which mess up the derivatives. To use indirect cues like shadows in geometric reconstructions, this issue must be fixed. (1/7)

Geometric reconstruction of a chair from several shadows.

wjakob,
@wjakob@fosstodon.org avatar

While much has been said and written about this problem, current methods often perform poorly. We present projective sampling, a new method to accelerate derivative evaluation by orders of magnitude. This is joint work with Ziyi Zhang and Nicolas Roussel. 2/7

wjakob,
@wjakob@fosstodon.org avatar

The key issue that projective sampling solves is to compute a Monte Carlo integral over boundary paths that graze the silhouette of an object. In a physically based renderer, these silhouettes are not static but depend on on the shading point. 3/7

wjakob,
@wjakob@fosstodon.org avatar

Not only is identifying silhouette boundaries a difficult task, most light paths that are on such a boundary are irrelevant. This space is incredibly sparse, which makes good sampling strategies that much more important. 4/7

video/mp4

wjakob,
@wjakob@fosstodon.org avatar

Our method relies on projecting area samples on a shape to its boundary. In fact, we can re-use the samples from the “primal” rendering. We can therefore efficiently build a guiding structure to sample boundaries in our scenes. 5/7

video/mp4

wjakob,
@wjakob@fosstodon.org avatar

The paper and video-talk for it are available here: http://rgl.epfl.ch/publications/Zhang2023Projective
The official implementation is already fully integrated in Mitsuba 3, if you wish to try it out for yourself! 6/7

turbobob, to science
@turbobob@mamot.fr avatar

#FOSDEM Could anyone recommend a useful resource for learning pybind11, in complement of the docs (https://pybind11.readthedocs.io/en/stable/index.html) :blobcatheart: ?

#Science #Python #Cpp #pybind11 #Physics #Research #Climate #FOSDEM24

wjakob,
@wjakob@fosstodon.org avatar

@turbobob the nanobind documentation ;) - https://nanobind.readthedocs.io/en/latest/

wjakob, to random
@wjakob@fosstodon.org avatar

I was promoted to associate professor 🥳.

https://actu.epfl.ch/news/appointment-of-epfl-professors-105/

wjakob, to random
@wjakob@fosstodon.org avatar

A response to a student who was surprised that I still bother working on "old" (non-AI/ML) topics in graphics.

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • anitta
  • InstantRegret
  • mdbf
  • osvaldo12
  • magazineikmin
  • tacticalgear
  • rosin
  • thenastyranch
  • Youngstown
  • Durango
  • slotface
  • everett
  • kavyap
  • DreamBathrooms
  • JUstTest
  • khanakhh
  • ethstaker
  • cubers
  • normalnudes
  • tester
  • GTA5RPClips
  • cisconetworking
  • ngwrru68w68
  • megavids
  • provamag3
  • Leos
  • modclub
  • lostlight
  • All magazines
    •