Fus3D: Decoding Consolidated 3D Geometry from Feed-forward Geometry Transformer Latents

2026-03-26

AI Summary

This work addresses the challenge of efficiently reconstructing complete 3D geometry from unstructured image collections by proposing a feed-forward approach that, for the first time, jointly extracts dense signed distance fields (SDFs) directly from intermediate features of a pre-trained multi-view geometric Transformer. A unified 3D implicit representation is constructed through voxelized embeddings combined with cross- and self-attention mechanisms, followed by a lightweight convolutional decoder to produce the full SDF. The method eliminates conventional per-view prediction and post-hoc fusion pipelines, instead introducing an occupancy-aware SDF supervision strategy that accommodates real-world scenarios such as non-watertight meshes. It achieves geometrically consistent and complete reconstructions under both sparse and dense input views, with inference times under three seconds—significantly outperforming existing fusion-based approaches.

Technology Category

Computer Vision: 3D Computer VisionMachine Learning: Multi-instance/Multi-view LearningKnowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphsSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We propose a feed-forward method for dense Signed Distance Field (SDF) regression from unstructured image collections in less than three seconds, without camera calibration or post-hoc fusion. Our key insight is that the intermediate feature space of pretrained multi-view feed-forward geometry transformers already encodes a powerful joint world representation; yet, existing pipelines discard it, routing features through per-view prediction heads before assembling 3D geometry post-hoc, which discards valuable completeness information and accumulates inaccuracies. We instead perform 3D extraction directly from geometry transformer features via learned volumetric extraction: voxelized canonical embeddings that progressively absorb multi-view geometry information through interleaved cross- and self-attention into a structured volumetric latent grid. A simple convolutional decoder then maps this grid to a dense SDF. We additionally propose a scalable, validity-aware supervision scheme directly using SDFs derived from depth maps or 3D assets, tackling practical issues like non-watertight meshes. Our approach yields complete and well-defined distance values across sparse- and dense-view settings and demonstrates geometrically plausible completions. Code and further material can be found at https://lorafib.github.io/fus3d.

Problem

• 3D reconstruction
• Signed Distance Field
• multi-view geometry
• feed-forward transformer
• dense geometry regression

Innovation

• Feed-forward Geometry Transformer
• Signed Distance Field (SDF)
• Volumetric Latent Grid
• Multi-view 3D Reconstruction
• Validity-aware Supervision