π3: PERMUTATION-EQUIVARIANT VISUAL GEOMETRY LEARNING

Technical Analysis "π3" is a feed-forward neural network that offers a novel approach to visual geometry reconstruction, breaking the reliance on a conventional fixed reference view. π3 employs a fully permutation-equivariant architecture to predict affine-invariant camera poses and scale-invariant local point maps without any reference frames. High-level Analogy: Imagine you and your friends are trying to build a 3D model of a new building, but none of you have a master blueprint or a single 'official' measuring stick. Instead, each of you takes pictures and notes down measurements from your own perspective, using your own rulers. The trick is, you all agree on how your rulers relate to each other (e.g., 'my ruler is twice as long as yours for this particular wall'), and you can figure out the exact position of each friend relative to the others. This new system, , is like everyone sharing their individual measurements and relative positions, without ever needing to pick one person's ruler or viewpoint as the 'master' one. This way, if someone's camera is shaky or their initial measurements are off, it doesn't mess up the whole group's understanding of the building. Everyone's contribution is equally important, and the final 3D model is robust, accurate, and doesn't care whose photo you look at first. Motivation of the Work Visual geometry reconstruction, the process of creating 3D models from images, has evolved from traditional methods like Structure-from-Motion (SfM) and…