Research Areas

Current Focus

SpatialLab is an independent research project working through spatial intelligence layer by layer: photorealistic 3D vision, Gaussian Splatting, neural rendering, and world models — from real rooms toward interactive worlds.

Project Overview

WHAT IS SPATIALLAB

Exploring urban reconstruction and interactive 3D environments.

3D City Replica Explorer

Exploring natural environments through immersive visualization.

Natural Spaces Monitor

Reconstructing and exploring interior spaces in 3D.

Indoor Mapping Replay

RECENT PROJECTS

LiDAR, photogrammetry, scene capture, open datasets

Spatial Data & 3D Models

3D rendering, spatial AI, visual understanding

Analysis & Visualization

Computer vision, neural rendering, world models

AI & ML

WebGL, Three.js, Unity, interactive experiences

Spatial Platforms

Open source tools and reproducible workflows

Open Infrastructure

Scene representations, datasets and interfaces

Data & APIs

RESEARCH STACK

从一个像素开始：从零构建可微 Volume Renderer

从零构建 NeRF：从多视图图像学习一个可渲染的三维场景

从零构建 3D Gaussian Splatting：用显式高斯学习一个可渲染的三维场景

友链

Github

友链副本

作品集

从一条射线出发，逐步构建三维 Density/Color Field，推导局部光传输、Transmittance 与 Volume Rendering Equation，并使用 PyTorch 从零实现单射线渲染、相机模型和完整图像渲染器。最后通过 Autograd 验证像素误差如何反向更新三维场参数，建立从三维表示到图像、再从图像反馈到三维表示的可微渲染闭环。

从上一篇可微 Volume Renderer 出发，使用已知相机参数的多视图 RGB 图像，从零构建一个完整的最小 NeRF。逐步实现 Positional Encoding、Stratified Sampling、Hierarchical Sampling 与 Coarse/Fine 训练，最终生成未见相机视角，并通过 RGB、Depth 和 Density 分析理解新视角合成与真实几何重建的区别。

承接 NeRF 篇的可微渲染与多视图场景学习，将隐式 Radiance Field 转换为可学习的显式 3D Gaussian 表示。逐步实现 Gaussian Shape、Perspective Projection、Alpha Compositing 和图像监督优化，并引入 SfM 初始化、Adaptive Density Control、Spherical Harmonics 与 Tile-Based Rasterization。最终完成新视角渲染与场景保存，通过 RGB、Depth、Geometry 和性能诊断理解高质量图像生成与准确三维重建的区别，为下一篇 RoomGS 真实室内重建奠定基础。

Spatial Intelligence From Scratch

友情链接

往期整理

历史归档

文章分类

文章标签

关于我

类型为Notice的文章将被显示为公告

公告

SEO Difficulty Score

External Mentions

Related Notion Pages

Content Freshness Tier

password

icon

date

type

CAN_COPY

slug

status

title

summary

Post Board

Config

Table

Post Gallery

From an Implicit Field to Gaussian Primitives

Building 3D Gaussian Splatting From Scratch

Chapter 1 — Rethinking Scene Representation

2.4 Learning Different Shapes in Different Directions

From Euclidean Distance to Mahalanobis Distance

Chapter 3 — Projecting Gaussians into Images

3.3 Why Perspective Breaks Gaussian Closure

Why Does a Distant Gaussian Look Smaller?

3.7 Building Our Gaussian Projection Module

How Accurate Is the Taylor Approximation?

Chapter 4 — Building Our First Splat Renderer

4.7 A Correct Renderer Is Not Yet a Fast Renderer

4.8 A Renderable Scene Is Not Yet a Learned Scene

5.1 From Forward Rendering to Inverse Graphics

5.2 Following the Gradient Through the Renderer

5.3 Is Our Renderer Really Differentiable?

Comparing Autograd with Finite Differences

5.5 When Correct Gradients Are Not Enough

Depth Sorting and Piecewise Differentiability

5.6 Did We Recover the Same 3D Gaussians?

5.7 Saving Our First Learned Gaussian Scene

Chapter 6 — Reconstructing a Scene from Images

6.2 Building a Reliable Multi-view Dataset

Initializing the Scale from Local Point Spacing

Initializing Rotation, Opacity, and Color

6.7 Diagnosing the Fixed-Capacity Baseline

Why Fixed Gaussian Count Becomes a Limitation