Implicit Neural Networks
约 129 个字
Occupancy network¶
.assets\image-20240225102149808.png)
In essence a classifier!
Representing Materials and Lighting¶
\(\Large{\mathbf L_{cSLF}(\mathbf{p,v,l}) : \mathbb R^3 \times \mathbb R^3 \times \mathbb R^M \rightarrow \mathbb R^3}\)
Representing Motion¶
Representing Scenes¶
Needs to exploit local features with CNN
.assets\image-20240225102729114.png)
Differentiable Volumetric Rendering¶
Learning from Images
No 3D supervisions
.assets\image-20240225102834653.png)
-
Forward
.assets\image-20240225102918368.png)
-
Backward
.assets\image-20240225102943720.png)
Neural Radiance Fields¶
Render image from novel viewpoints, instead of fine 3D reconstruction
.assets\image-20240225103126139.png)
-
Fourier Features
Learn high dimensional function in low dimensional space
Avoid over-smoothing, more sharp details
Generative Radiance Fields¶
Can train from unstructured and unposed image collections. (With no camera pose)
.assets\image-20240225103459433.png)
.assets\image-20240225103516244.png)
View dependent appearance and view independent density! Disentangle appearance and occupancy.
-
GIRAFFE
.assets\image-20240225103631222.png)
Different parts of an image
Wrap up¶
.assets\image-20240225103657686.png)