Gaussian Splatting: 3D Reconstruction and Novel View Synthesis, a Review(1)

ABSTRACT

Image-based 3D reconstruction presents a complex challenge requiring the inference of an object's or scene's three-dimensional shape from a collection of input images. Learning-based approaches have garnered considerable attention due to their capability to directly estimate 3D structures. Such review papers focus particularly on state-of-the-art techniques for 3D reconstruction, with an emphasis on the creation of novel, unseen perspectives.

Figure

Figure 1

Estimated figure of publications pertaining to Gaussian Splatting since its establishment in June 2023.

Figure 2

Traditional 3D data representations.

Figure 3

Novel 3D data representations.

Figure 4

3D Gaussian Splatting architecture.

Figure 5

Dynamic and deformation based methods.

Figure 6

Dream-Gaussian framework generating images through iteration.

Figure 7

Lucid Dreamer framework with iterative navigation and dreaming

Figure 8

The D3GA system, starting from the left side, includes: joint angles; predicted body structure; predicted upper and lower cages; three-dimensional Gaussian distributions; clothing components; and final output image.

Figure 9

4D Gaussian rendering from HiFi4G.

Figure 10

HeadGaS framework generating realistic head avatars.

Figure 11

GS-SLAM Framework.

Figure 12

PhysGaussian Framework training.

Figure 13

GaussianEditor changing scenes with various text prompts