For those interested in normalized gradient methods and optimal transport: I introduce a new class of "spectral" Wasserstein distances for which spectrally normalized gradient descent (Muon but without momentum and small step size ...) is a spectral-W gradient flow: arxiv.org/abs/2604.04891
Could you elaborate please? If the reviews remain anonymous, how making them public could improve their quality? By hoping that the reviewers would feel more pressure? I'd be curious to see if this policy has improved the quality of reviews where it has been used🤔
I relate to this so much! Luckily enough my advisor is here to filter what is really relevant to the reader, but it's always a pain in the heart to let go of the remarks that you felt gave more depth🥲we should have a "director's cut" document to add all the stuff we were the only ones to care about😂
I have procrastinated on writing so much that I wrote an entire document on writing tips: cseweb.ucsd.edu/~tzli/writin...
Probably not much is new but I find I still need to repeat the same things to my students regularly. Will update this document over time hopefully.
Exactly! You're welcome, thank you again for your kind words🙏
Thank you so much🙏 on second thoughts if I remember well I think that I tried the joint PCA at the start of the project and that it did not work well... Also the problem with only taking the first eigenvector is that the BSP matching would not be random hence no merging...
Hi! Just to make sure, the first principal component is not used here it's just the common heuristic for building a standard BSP for rendering. I did not try with the concatenation but I feel like it's not a good approach since it is not invariant to a translation to one of the point clouds.
Thank you!
Thank you so much for your kind words once again ! It means the world🙏🙏🙏
I am very lucky to work with @nbonneel.bsky.social Vincent Nivoliers and @dcoeurjo.bsky.social !
Super happy and honored to share that our paper "BSP-OT: Sparse transport plans between discrete measures in log-linear time" won a *Best paper award* at SIGGRAPH Asia 2025!
If you are here, come see my presentation about this work Wednesday afternoon!
Many thanks to the award committee!
The Graphics Replicability Stamp Initiative (GRSI, www.replicabilitystamp.org), a community-driven initiative to promote replicability in Graphics research, is seeking volunteers.
More details in the 'Volunteering' section of the home page.
I could try between points embeded in high dimensions with the spectral embedding we use to sample meshes for instance
We had experiments using the toy datasets of some ML papers (between gaussians for example) in dimension up to 20 (and it behave well!) but we mainly lacked interesting data, if you have ideas or references to get some in that kind of dimensions that would be great 🙏
Of course! That would be awesome!
The code interface is very simple (at least in the bijective case).
We mainly tested BSP-OT on graphics applications in low dimension, I hope that POT users would still find it usefull !
Happy to share the excellent work of @baptiste-genest.bsky.social on making sliced optimal transport *much* better by replacing the sort along a line with a variant of QuickSort that directly takes multiples directions at once! It also directly provides a coupling between points.
Glad to share with you our last paper to be presented @ SIGGRAPH Asia 2025 !
Kudos @baptiste-genest.bsky.social
Check the paper : baptiste-genest.github.io/papers/bsp-o...
Code : github.com/baptiste-gen...
I will present this work during the conference! (15-18 december)
bijective shape interpolation, comparaison with geomloss
Intrinsic manifold sampling
color transfert comparison
partial point cloud registration
We show that this approach can be applied to standard applications of OT, like image stippling, shape interpolation, color transfer, partial point cloud registration, intrinsic mesh sampling and Topological Data Analysis (a domain where bijectivity is required!).
local assignment swapping, only consider 4 points
global merging : select best assignment on each connected component of the union
We also extend it to partial matchings, and general sparse transport plans between non-uniform measures.
Then, in the bijective and injective cases, we exploit this extreme sparsity to efficiently merge two assignments using very simple discrete optimization.
Remarkably, this procedure is also a generalization of the (projection + 1D assignment) step used at the core of Sliced OT! The difference being that it preserves much more information than a single projection, which yields couplings of much higher quality, with the same complexity.
BSP matching : recursive splitting of the points along linear slices, hence matching the leaves of two BSPs
BSP Matchings generalize the Quicksort algorithm. We simultaneously build two BSP trees on each point cloud and match their leaves.
paper teaser
This is a joint work with @nbonneel.bsky.social, Vincent Nivoliers and @dcoeurjo.bsky.social @cnrsinformatics.bsky.social
Our pipeline is simple: we use BSP (Binary Space Partitioning) matchings to efficiently generate random bijections before merging them into a single one of low transport cost.
Computing the exact bijection of the optimal transport (OT) problem between very large point sets is completely untractable…
In our SIGGRAPH Asia 2025 paper: “BSP-OT: Sparse transport plans between discrete measures in log-linear time” we get one with typically 1% of error in a few seconds on CPU!
If N is a Binomial random variable with parameters n and p, then the expected value of "N choose k" is "n choose k" time p raised to the k-th power.
"The greatest trick the Binomial ever pulled was convincing the world it didn't behave exactly as expected."
Carte de Paris simulation îlot de chaleur
Vous avez sûrement déjà lu que la climatisation aggrave les canicules car l’air chaud rejeté à l’extérieur augmenterait la température des villes de 2 à 3°C.
Vérifions ce que disent vraiment les études scientifiques à ce sujet 🧑🔬
(lisez jusqu’au bout, car vous allez être 🤯🤯 à la fin)
#Thread
⬇️⬇️
📜 New SIGGRAPH 2025 paper 🎉
❔How to compute bounding volumes for procedural Signed Distance Fields (SDFs)? This is not so trivial!
💡We propose a simple method called Sphere Carving. It extracts (convex) bounding volumes around SDFs, requires very few evaluations, and is GPU compatible.
OPEN SCIENCE IS THE FUTURE!
Very pleased to see the Association for Computing Machinery (ACM — the parent organization to my “home” conference SIGGRAPH) has announced it is transitioning to full open access by January 2026!!
#ACM #open #access #science #publication #journal #compsci
I am glad to attend #HPG2025 where I just presented our paper "Fused Collapsing for Wide BVH Construction", co-authored with Mathias Paulin. We propose a fast build algorithm for wide BVHs that directly computes a wide hierarchy without additional collapsing pass. Webpage: wbrbr.org/publications...
Thank you so much for your feedback 🤗🤗🤗 I was very lucky, the ideas were very simple and worked very quickly. From the first ideas to the submission it took around 4 months and a half🙏
