diff --git a/index.html b/index.html index 48ccebc..49d8877 100644 --- a/index.html +++ b/index.html @@ -63,12 +63,12 @@
-
+
-
+
-

Generative Densification

-

Learning to Densify Gaussians for High-Fidelity Generalizable 3D Reconstruction

+

Generative Densification: Learning to Densify Gaussians

+

for High-Fidelity Generalizable 3D Reconstruction

@@ -144,7 +144,7 @@

Learning to Densify Gaussians for High-

Our method selectively densifies (a) coarse Gaussians from generalized feed-forward models. (c) The top K Gaussians with large view-space positional gradients are selected, and (d-e) their fine Gaussians are generated in each layer. - (g) The final Gaussians are obtained by combining (b) the remaining (non-selected) Gaussians with (f) the union of each layer's Gaussians. + (g) The final Gaussians are obtained by combining (b) the remaining (non-selected) Gaussians with (f) the union of each layer's output Gaussians.

@@ -158,7 +158,7 @@

Abstract

-

+

Generalized feed-forward Gaussian models have achieved significant progress in sparse-view 3D reconstruction by leveraging prior knowledge from large multi-view datasets. However, these models often struggle to represent high-frequency details due to the limited number of Gaussians. While the densification strategy used in per-scene 3D Gaussian splatting (3D-GS) optimization can be adapted to the feed-forward models, it may not be ideally suited for generalized scenarios. @@ -173,20 +173,23 @@

Abstract

- +
-

Method Overview

+

Method

+ + +

Generative Densification Overview

-

+

Our method selectively densifies coarse Gaussians generated by feed-forward Gaussian models: -
    -
  1. +
      +
    1. The top K coarse Gaussians with large view-space positional gradients are selected.
      2. -
    2. +
    3. The selected Gaussians are processed through the densification module to generate fine Gaussians.
    4. @@ -196,9 +199,23 @@

+ + +

Object-level & Scene-level Pipelines

+
+
+ +

+ We present two models incorporating our densification method, based on LaRa and MVSplat. + For object-level reconstruction, fine Gaussians are generated using the Gaussians and volume features produced by the LaRa backbone (top row). + For scene-level reconstruction, fine Gaussians are generated per view by utilizing the pixel-aligned Gaussians and image features extracted from the MVSplat backbone (bottom row). +

+
+
+
- +
@@ -206,8 +223,131 @@

Object-level Reconstruction

+

Quantitative Comparisons

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Method#Param(M)GobjaverseGSOCo3D
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
MVSNeRF0.5214.480.8960.18515.210.9120.15412.940.8410.241
MuRF15.714.050.8770.30112.890.8850.27911.600.8150.393
LGM41519.670.8670.15723.670.9170.06313.810.7390.414
GS-LRM300---30.520.9520.050---
LaRa12527.490.9380.09329.700.9590.06021.180.8620.216
Ours13428.580.9450.08031.060.9660.05821.720.8650.209
Ours (w/ residual)13428.750.9460.07831.230.9670.05822.080.8670.206
 + Quantitative comparison results. + Our model achieves the highest PSNR in both in-domain reconstruction and cross-dataset generalization tasks, + even outperforming GS-LRM (the current state-of-the-art model for object-level reconstruction). +
+ -

Comparisons against LaRa

+

Qualitative Comparisons