[7/7]
For more interactive visualizations, check out our project page: https://hanyuc.com/archsym
Huge thanks to the team! Ruojin Cai (@ruojin8), Steve Marschner, and Noah Snavely (@Jimantha).
Come by our poster on Sunday, Poster Session 6, if you’re interested!
[6/7]
On held-out landmark scenes, our method improves over both Reflect3D (finetuned on our dataset), and a simple baseline that directly regresses plane parameters from the same VGGT features.
See our paper for full quantitative results and evaluation details.
[3/7]
While a failure mode for SfM, these “illusory” matches are the exact correspondences needed to define a building's symmetry.
To find reflectional symmetries, we simply *flip* one image before matching. We lift matches to 3D, fit planes, and cluster them into annotations.
[2/7]
Moving into the wild raises a practical question: how do we find & annotate real symmetric structures?
Architectural landmarks are a perfect starting point, and the “doppelganger” problem gives us a clue: image matchers easily confuse visually similar building structures.
[4/7]
We then build a symmetry detector based on VGGT, a powerful 3D foundation model.
From a single image, scene scale is ambiguous. So instead of predicting a plane in some arbitrary coordinate system, we ground the symmetry plane relative to VGGT’s predicted point map.
[5/7]
Specifically, we predict a signed distance for each 3D point—indicating which side of the plane it lies on, and how far away.
This turns plane prediction into a dense prediction problem, allowing us to more accurately anchor the symmetry plane to the scene geometry.