When parking, you can now see a high fidelity 3D representation of the world around your vehicle, including proximity & shape of nearby objects, barriers, vehicles & painted road markings
By using a dedicated neural network to model obstacles & paint lines, we can accurately
NeRFactor is out! It's a physically-based model that factorizes appearance into shape and reflectance given just multi-view images under *one unknown* illumination (i.e., NeRF data). It supports free-viewpoint relighting (w/ shadows!) and material editing.
🎉NeRFactor has been conditionally accepted to SIGGRAPH Asia 2021! Looking forward to presenting and chatting about it this December in Tokyo or online.
NeRFactor is out! It's a physically-based model that factorizes appearance into shape and reflectance given just multi-view images under *one unknown* illumination (i.e., NeRF data). It supports free-viewpoint relighting (w/ shadows!) and material editing.
📢We (Marc Levoy's computational photography team at Adobe) are looking to hire research interns for Summer 2022. Ping me if interested in working with us on relighting, neural rendering, or any computational photography problem. (Please help RT!) More info, in Marc's own words:
Check out our latest work, Neural Light Transport, on *simultaneous* relighting💡 and view synthesis 📷 by learning to interpolate a 6D light transport function in the texture space!
Paper:
Video:
Project:
Introducing "Neural Light Transport": Embedding a convnet within a predefined texture atlas enables *simultaneous* view synthesis and relighting, while maintaining backwards compatibility with oldschool graphics engines. Great work
@xiuming_zhang
! More at
Hot take after reviewing for
#CVPR2023
: IMO, we should stop calling our own methods "novel." It is subjective and carries no meaning in a scientific publication. When I was submitting to science journals, the editorial office checked for such words and asked me to remove them👍.
Speaking of "novel," I stopped, many years ago, using novelty as a criterion to judge papers. What's novel, and what's not? Is it a subjective or objective call? NeRF is amazing, but volume rendering, MLPs, and PE have been around. More important is whether the method works, IMO.
Wanna build AI autonomy that runs in the real world? Come talk to us at
@Tesla
AI! You’ll find exciting problems to work on whether it be cars
#autopilot
or robots
#optimus
, and whether you be a high-level or mid-level vision person!
#CVPR2023
@Tesla
AI team is at
@CVPR
in Vancouver this week! If you are also here, stop by and check out what we have been working on for Autopilot, Optimus, and dojo!
#CVPR2023
Ever feeling NeRVous about tracing to every single light plus indirect illumination? NeRV can come to your rescue! It jointly optimizes for shape, light visibility, reflectance, & indirect illumination (!), while staying still tractable.
Our code on editing conditional radiance fields is out! We can edit the shape and color of 3D regions with simple user scribbles. With
@xiuming_zhang
, Zhoutong Zhang,
@rzhang88
,
@junyanz89
, Bryan Russell.
Paper + Code + Video + Demo:
Wanna stay updated on what your favorite researchers are up to?🔔Check out , which compares the webpages' current HTMLs against their previous snapshots and sends you an email of the deltas. Great for tracking folks' new papers, positions, etc. PRs welcome!
Can you make a jigsaw puzzle with two different solutions? Or an image that changes appearance when flipped?
We can do that, and a lot more, by using diffusion models to generate optical illusions!
Continue reading for more illusions and method details 🧵
Cool! "Preconditioner" sounds like camera parameter-dependent scales that "normalize" magnitudes of different camera parameters' effects so that all types of parameters receive more equal gradients during optimization. Keunhong, I would've guessed you'd name this "Equalizer." 😉
Introducing CamP🏕️ — a method to precondition camera optimization for NeRFs to significantly improve quality. With CamP we’re able to create high quality reconstructions even when input poses are bad.
Project page:
ArXiv:
(1/n)
Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis
We model the world as a set of 3D Gaussians that move & rotate over time. This extends Gaussian Splatting to dynamic scenes, with accurate novel-view synthesis and dense 3D trajectories.
📢Consider submitting your work to the Technical Communications & Posters programs of SIGGRAPH Asia 2021, which is still scheduled to be a physical event in Tokyo🗼, with the option of online presentation if preferred -- great regardless of your travel preference!
in addition to generating videos from text, Sora can morph between two videos. here's one example I love where it starts as a drone and turns into a butterfly
Check out our new paper that turns a (single image) => (interactive dynamic scene)!
I’ve had so much fun playing around with this demo.
Try it out yourself on the website:
JaxNeRF! Today we're releasing Google's internal JAX implementation of NeRF. Training goes from 3 days to 2.5 hours (on a TPU pod), PSNR is slightly higher(?!), and it has all the functional/gradient goodness we love about JAX. Great work
@a_k_a_Billy
!
I voted nay to this motion too because I think it's awkward to post your new paper to arXiv and meanwhile have to stay quiet about it. The ban also sounds hard to implement in practice (e.g., bot posting/tweeting is ok), although I guess most researchers will just abide by it.
Social media is important to share research publicly and for junior researchers it's one of few ways to get yourself known.
So it's not a great move to take that away when conferences are mostly virtual - I believe the CVPR motion was a big mistake and should be re-considered.
@jon_barron
@dimadamen
Spot on! People tend to call any model comprising components originally devised for real generative tasks (like diffusion models) generative. Diffusion model-based monocular depth estimation: generating depth using diffusion models (conditioning on RGB) -> generative. 😂
@jon_barron
"Looping over pixels vs. primitives" -- insightful! +1 to the "neural" comment, too. Been viewing NeRF as optimization instead of learning. Then, is it, ahem, "AI"? General public probably thinks it is. Meta-NeRF sounds more like AI than vanilla NeRF. Regardless, NeRF is cool!
@elliottszwu
If you set the check frequency high enough (like every minute), you can indeed catch a fellow researcher fiddling with their webpage as it happens! 🕵️♂️👁️👀
@taiyasaki
I've been having this urge to defer it to the final section or to right before "Conclusion" for the same reasons you listed, but never ended up doing it, partly to avoid upsetting "traditional" people (who may happen to be my reviewers).