FreeSplat: Generalizable 3D Gaussian Splatting Towards Free-View Synthesis of Indoor Scenes

TL;DR

A novel framework FreeSplat is presented that is capable of reconstructing geometrically consistent 3D scenes from long sequence input towards free-view synthesis and performs inference more efficiently and can effectively reduce redundant Gaussians, offering the possibility of feed-forward large scene reconstruction without depth priors.

Abstract

Empowering 3D Gaussian Splatting with generalization ability is appealing. However, existing generalizable 3D Gaussian Splatting methods are largely confined to narrow-range interpolation between stereo images due to their heavy backbones, thus lacking the ability to accurately localize 3D Gaussian and support free-view synthesis across wide view range. In this paper, we present a novel framework FreeSplat that is capable of reconstructing geometrically consistent 3D scenes from long sequence input towards free-view synthesis.Specifically, we firstly introduce Low-cost Cross-View Aggregation achieved by constructing adaptive cost volumes among nearby views and aggregating features using a multi-scale structure. Subsequently, we present the Pixel-wise Triplet Fusion to eliminate redundancy of 3D Gaussians in overlapping view regions and to aggregate features observed across multiple views. Additionally, we propose a simple but effective free-view training strategy that ensures robust view synthesis across broader view range regardless of the number of views. Our empirical results demonstrate state-of-the-art novel view synthesis peformances in both novel view rendered color maps quality and depth maps accuracy across different numbers of input views. We also show that FreeSplat performs inference more efficiently and can effectively reduce redundant Gaussians, offering the possibility of feed-forward large scene reconstruction without depth priors.

Figures (10)

• Figure 1: Comparison between FreeSplat and previous methods. pixelSplat pixelsplat and MVSplat mvsplat fail to reconstruct geometrically consistent global 3D Gaussians, while our FreeSplat is proposed to accurately localize 3D Gaussians from long sequence input and support free view synthesis.
• Figure 2: Framework of FreeSplat. Given input sparse sequence of images, we construct cost volumes between nearby views and predict depth maps and corresponding feature maps, followed by unprojection to Gaussian triplets with 3D positions. We then propose Pixel-aligned Triplet Fusion (PTF) module, where we progressively aggregate and update local/global Gaussian triplets based on pixel-wise alignment. The global Gaussian triplets can be later decoded into Gaussian parameters.
• Figure 3: Visual illustration of PTF. The PTF incrementally projects current global Gaussians to input views and computes their pixel-wise distance with local Gaussians. Nearby local Gaussians are then fused using a lightweight Gate Recurrent Unit (GRU) network gru.
• Figure 4: Qualitative Results of Long Sequence Explicit Reconstruction. For each sequence, the first two rows are view interpolation results, and the last two rows are view extrapolation results.
• Figure 5: Qualtitative Ablation Study. The first and second row use input view lengths of 3 and 10.