What are we focusing on?
- Predicting 3D shapes from sigle image.
- Predicting labels from 3D shapes, such as auto-drive(using depth imformation).
Topics
- Computing correspondences
- SLAM
- Self-supervised learning
- Structure from motion
- Multi-view stereo
3D shape representations
- Depth map RGB+depth
- Voxel Grid
- Pointcloud
- Mesh, using in CG
- Implicit Surface, Nerf(using a function to represent the 3D shapes)
Depth map
RGB image+Depth map = RGB-D Image, namely , it has been used as 2.5D.
Only recoed the relative distance, not exactly distance.
This type of data can be recorded directly for some tyoes of 3D sensors, otherwise, it can be predicted by Neural Network easily.
Date:2014/2015, using fully conv nevwork, and it can be predicted easily.WHY, first ,the model shoule be able to recognize the object first.
Problem: Scale/Depth Ambiguity:Small&clase–Large&far
BAD PERFOMEANCE? ADD LOSS, in this instance, we add scale invariant loss. We predict the surface normals(if the pixels share the same surface normal, they are included in the same surface). But it will have a bad performance if the illumination is not even.
Voxels
- Represent a shape with a V* V * V grid of occupancies. Fill the grid with voxels.
- Just like segmentation masks in Mask R-CNN
- Disadvantages: no details & noTexture
- Tow topics:Generative Voxel Shapes / 3D classification
- Disadvantages: Memory, more lossing details.
Improve: Octree
3D Pointcloud
Name one question: WHY 3D? to edit the shapes in 3D, so the discrete representation is unique in 3D tasks.
- Can represent fine structures without huge numbers of points.
- Requires new model architecture and new losses, etc.
- Doesn’t explictly represent the surface of the shapes, no 3D features. extracting a mesh for rendering.
- Point net.
- not many Pointcloud Generation, but focusing on the understanding of pointcloud
Mesh
All 3D representation will be rendered to the mesh, the final form of the 3D model.
Key Concepts: Vertices/ Faes
Advantage: Standard+Explicitly+Adaptive(allocate more faces to areas with fine
detail)
How to represent texture? U-V Map.[another topic, no need to be mentioned]
Method: Pixel2Mesh
- Iterative mesh refinement(before my evl, I don’t know how many vertives to use)
Implicit Surface
Implicit function//Explicit shape
Nerf: generate images from different perspectives.