3DL is recruiting PhD students to start in Fall 2025!
A thread👇highlighting some of our recent works, and what we are excited to explore next 🚀
Apply by Dec. 16; more info: https://t.co/WyvimLwjXe 1/
Mila's annual supervision request process opens on October 15 to receive MSc and PhD applications for Fall 2025 admission! Join our community! More information here https://t.co/r01eLcXtZw
@sellan_s I think the implicit function is indeed a function: the (local) function that is represented by the zero level set (e.g., in the case of a surface, the *function* from a 2D patch to 3D). That is the "implicit function" that the "implicit function theorem" refers to, methinks.
@amirvaxman_dgp For one, I can say that I've never been able to use the LBFGS solver in Pytorch - always stagnates on problems that I know would've worked with Matlab's LBFGS solver
Interested in joining Mila's research community? Our annual supervision request process for new Mila students is starting this Sunday October 15, 2023.
For more information ➡️ https://t.co/Z490dtJQGD
Excited to be presenting DA-Wand: Distortion Aware Selection Using Neural Mesh Parameterization at #CVPR2023!
Come find us at the Thursday AM session (West Hall ABC)!
Poster ID: THU-AM-025
Project website: https://t.co/LXbAomU7og
Paper link: https://t.co/Ds8EGDGkk2
(1/N)
@amirvaxman_dgp No guarantees about prediction in the triangle's (irrelevant) normal direction. It's a null-space for the loss, so intuitively, probably the network chooses its prediction in the normal direction s.t. it facilitates making a correct prediction in the tangent space.
Excited to finally share *Neural Jacobian Fields* - our SIGGRAPH 2022 paper on learning highly-accurate deformations and mappings of 3D meshes, in a triangulation-agnostic manner. https://t.co/EpFqzYbLGw
(1/10)
@amirvaxman_dgp The loss is defined on the restriction, and on it it is well-defined. The *gradient* is also well-defined; it's simply zero in each triangles' normal direction (that's due to back-propagation through the restriction operator which discards the normal direction).
@esx2ve Yes, any genus, out of the box! More so, your dataset isn't required to have a consistent genus. (Fig. 9 shows how a network trained solely on genus-0 meshes still makes valid predictions for higher-genus meshes.)
@coreqode We learn to emulate SLIM's locally-injective maps. Neither us nor SLIM have guarantees for global bijectivity, but this could be achieved by training with maps that were generated via a globally-bijective method (e.g., SLIM + ensuring non overlapping boundary)
I am personally very excited about this work. Mapping meshes is central to the graphics pipeline (e.g., modeling, deformations, UV maps), and hopefully this paper will open up new possibilities for seamlessly integrating machine-learning into real world 3D applications.
(9/10)