3D Scene Analysis¶

Note

start on: Aug 23, 2023

Principal of Indoor Scene Network Design¶

From the talk of Richard in 2023/9/3 - 2023/9/4

consistency
- should be considered frist
- weak prior: same class = same function
- e.g. cycle consistency (for unpaired domain translation)
- e.g. VAE (predictable & compressibility \(\to\) need redundancy [symmetry] )
  - bottle neck layer in neural network: force to compress
compactness
- model representation
continuity
- linkage between the scene (bridge gap)

scene is difficult to align
scene has much more variations
real scenes are much messy
- note that: messy \(\neq\) random
- why messy: human action & activity (function of objects & co-occurrence of objects)
sub-scene indicate functional units

Problem #1: compare complex scene & layout
- In scene we need focal points (= subscenes)
- Find focal points (can be regarded as attention machoism in NLP)
  - context will make it easier
  - clustering algorithm [sig 2014]
  - find frequent & discriminate parts
  - do organize or retrieval
Problem #2: action-driven scene evolution
- should consider not only see/observe
- should consider functionality & human action
- scene understanding = inference of human action (annotated photo \(\to\) action model \(\to\) action graph)
Problem #3: text-driven 3D scene synthesis
- using scene proxy to guide
Problem #4: generative model of 3D scene structure

signed distance from the edge of a shape
- inside shape: negative
- outside shape: positive
- exactly on the shape: 0

游戏编程挑战：生成无边的水下世界体绘制算法 | Coding Adventure: Marching Cubes

使用 3D 空间中的一个点作为输入，随后返回一个值

\[ f(x,y,z) \to v \]

假设在空间中的某个区域内，我们可以使用函数 \(f\) 等间距的采样一些点

不妨设函数 \(f\) 的最大值为 \(16\)，最小值为 \(-32\)，如果存在一个表面阈值 \(sf\)，当表面阈值从 \(-32\) 增长到 \(16\) 时，均匀采样点就会逐渐消失（视频 0:45）

我们可以认为消失的点是空白的空间，而大于等于表面阈值的点，则处于形体的表面或者内部

如果把这个问题简化到一个立方体内部，如果白色的点位于物体内部，它就被激活，这样我们就得到了一个三角面片

Last update: September 12, 2023
Created: August 24, 2023