Wild3R: Feed-Forward 3D Gaussian Splatting from Unconstrained Sparse Photo Collection

2026-06-10

AI Summary

Existing feed-forward 3D Gaussian splatting methods struggle to robustly handle highly variable illumination and transient objects in real-world sparse photo collections. To address this challenge, this work proposes Wild3R, a feed-forward 3D Gaussian splatting approach tailored for unconstrained sparse image sets. The key innovations include the construction of WildCity—a large-scale dataset encompassing 200 scenes with 170 distinct illumination and transient object configurations—the design of a reference-view-conditioned feed-forward network to model cross-view appearance consistency, and the integration of a transient object removal mechanism. Experiments demonstrate that Wild3R significantly outperforms existing feed-forward methods across multiple metrics, achieving reconstruction quality comparable to traditional per-scene optimization approaches while substantially improving sparse-view 3D reconstruction in complex real-world scenarios.

Abstract

Feed-forward 3D Gaussian Splatting (3DGS) removes the need for time-consuming per-scene optimization required by traditional 3DGS. However, existing feed-forward approaches struggle with real-world photo collections that include diverse lighting conditions and transient objects. In this paper, we present Wild3R, a feed-forward approach for unconstrained sparse photo collections. The main bottleneck is the lack of training data that provides multiple viewpoints, a variety of illuminations, and transient variations necessary for learning robust scene representations. To address this, we introduce the WildCity dataset, which comprises 200 scenes, 170 lighting conditions, and transient objects, resulting in 337,500 images in total. By leveraging the dataset, our model learns appearance consistency across viewpoints conditioned on reference views, while removing transient content. Extensive experiments demonstrate that our method outperforms existing feed-forward approaches and achieves results competitive with prior per-scene optimization-based methods.

Problem

• Feed-forward 3D Gaussian Splatting
• unconstrained sparse photo collection
• transient objects
• illumination variation
• multi-view consistency

Innovation

• Feed-Forward 3D Gaussian Splatting
• Unconstrained Photo Collections
• Transient Object Removal
• Appearance Consistency
• WildCity Dataset