$π^3$: Scalable Permutation-Equivariant Visual Geometry Learning

2025-07-17

AI Summary

Traditional visual geometric reconstruction relies on a fixed reference view, leading to instability or failure when the reference is poorly chosen. To address this, we propose the first fully permutation-equivariant framework for visual geometric reconstruction—eliminating reference-view dependency entirely. Our method employs a permutation-equivariant neural network that jointly predicts affine-invariant camera poses and scale-invariant local point maps, enabling robust, input-order-agnostic modeling. The architecture is a feedforward neural network with rigorously enforced permutation equivariance, guaranteeing symmetric handling of arbitrary input image permutations. We validate our approach on three core tasks: monocular/video depth estimation, camera pose estimation, and dense point map reconstruction. It achieves state-of-the-art performance across all benchmarks, demonstrating substantial improvements in generalization, robustness to input ordering and occlusion, and scalability to varying numbers of input views.

Technology Category

• Computer Vision: Representation Learning for Vision
• Machine Learning: Learning with Manifolds
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We introduce $π^3$, a feed-forward neural network that offers a novel approach to visual geometry reconstruction, breaking the reliance on a conventional fixed reference view. Previous methods often anchor their reconstructions to a designated viewpoint, an inductive bias that can lead to instability and failures if the reference is suboptimal. In contrast, $π^3$ employs a fully permutation-equivariant architecture to predict affine-invariant camera poses and scale-invariant local point maps without any reference frames. This design makes our model inherently robust to input ordering and highly scalable. These advantages enable our simple and bias-free approach to achieve state-of-the-art performance on a wide range of tasks, including camera pose estimation, monocular/video depth estimation, and dense point map reconstruction. Code and models are publicly available.

Problem

• Breaks reliance on fixed reference view in geometry reconstruction
• Predicts affine-invariant camera poses without reference frames
• Achieves robust performance in diverse visual geometry tasks

Innovation

• Permutation-equivariant architecture for geometry learning
• Predicts affine-invariant camera poses without reference
• Scale-invariant local point maps for robust reconstruction