Project 5: Cindy Wei

README.md

@@ -2,25 +2,76 @@
 
 **University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 5**
 
-* (TODO) YOUR NAME HERE
-* Tested on: (TODO) **Google Chrome 222.2** on
-  Windows 22, i7-2222 @ 2.22GHz 22GB, GTX 222 222MB (Moore 2222 Lab)
+* Cindy Wei
+* Tested on: **Google Chrome 128.0** on
+  Windows 11, i7-13700K @ 5.4GHz 32GB, RTX 4080 16GB (Personal Machine)
 
 ### Live Demo
+https://chapiniwei50.github.io/Project5-WebGPU-Gaussian-Splat-Viewer/
+### Demo Video/GIF
 
-[![](img/thumb.png)](http://TODO.github.io/Project4-WebGPU-Forward-Plus-and-Clustered-Deferred)
+https://youtu.be/IgWtlDe2ayY
 
-### Demo Video/GIF
+## Project Description
+This project implements a real-time 3D Gaussian Splat Viewer using WebGPU, based on the 3D Gaussian Splatting technique. The viewer loads and renders pre-trained Gaussian scene files with sorting and basic rendering. **Note: We were unable to achieve smooth surfaces and proper blending - the implementation currently renders as discrete particles rather than continuous surfaces.**
+
+## Features Implemented
+
+### Part 1: Point Cloud Renderer
+- MVP calculation in vertex shader
+- Point cloud visualization
+
+### Part 2: Gaussian Splat Renderer
+- **Preprocessing Compute Shader**:
+  - View frustum culling with 1.2x bounding box
+  - 3D covariance computation from rotation and scale
+  - 2D conic and quad size calculation in NDC space
+  - Spherical harmonics color evaluation
+  - Storage of 2D Gaussian data for rasterization
+
+- **Rendering Pipeline**:
+  - Depth-based sorting using radix sort
+  - Indirect draw calls with dynamic instance count
+  - Quad reconstruction in vertex shader
+  - Basic fragment shader implementation
+  - **Implementation Limitation**: Failed to achieve proper conic-based splat evaluation and smooth surface blending
+
+## Screenshots
+
+**Point Cloud**
+![Point Cloud Rendering](images/pointcloud.png)
+
+**Gaussian**
+![Gaussian Splat Rendering](images/gaussian.png)
+
+## Performance Analysis
+
+### 1. Renderer Comparison
+- **272K Gaussians**: Point Cloud ~60 FPS, Gaussian Renderer 58-61 FPS
+- **1.06M Gaussians**: Point Cloud ~60 FPS, Gaussian Renderer 14-19 FPS
+- Gaussian rendering provides color but at significant performance cost with larger scenes
 
-[![](img/video.mp4)](TODO)
+### 2. Gaussian Count Scaling
+Performance scales inversely with Gaussian count:
+- **272K Gaussians**: 58-61 FPS
+- **1.06M Gaussians**: 14-19 FPS
+- 4x more Gaussians results in 3-4x performance drop
 
-### (TODO: Your README)
+### 3. View-Frustum Culling Effectiveness
+- **Bonsai Scene**: 96.6% of Gaussians rendered (3.4% culled)
+- **Bicycle Scene**: 100% of Gaussians rendered (0% culled)
+- Current culling implementation is largely ineffective
 
-*DO NOT* leave the README to the last minute! It is a crucial part of the
-project, and we will not be able to grade you without a good README.
+### 4. Performance Characteristics
+- **Bonsai (272K)**: Average frame time 16-17ms
+- **Bicycle (1.06M)**: Average frame time 60-62ms
+- Major bottleneck is Gaussian processing and sorting overhead
 
-This assignment has a considerable amount of performance analysis compared
-to implementation work. Complete the implementation early to leave time!
+## Technical Challenges
+- **Conic Evaluation**: Unable to properly implement the conic matrix evaluation in the fragment shader
+- **Blending Pipeline**: Challenges with alpha blending and proper transparency sorting
+- **Surface Continuity**: Failed to achieve smooth surface reconstruction from individual splats
+- **Culling Optimization**: View-frustum culling requires improvement for better performance
 
 ### Credits