Sehyeon Park
Back to projects
May 27, 2025
3 min read

StreamSplat

Streams 3DGS scenes in the browser with color/depth data and depth-based Gaussian-mesh fusion.
See Demo
KIST
Period
2025.03–2026.04
My role
First author · system architecture / implementation
Work

Server 3DGS/4DGS renderer and browser WebGL renderer separation

RGB-D/camera interface, WebCodecs decode, WebGL depth fusion

Client-only vs. hybrid streaming evaluation; Web3D paper writing

Outcome

Web3D 2025 Best Paper

Overview

StreamSplat is a web-based neural graphics system for streaming reconstructed 3DGS scenes and composing them with browser-side mesh content. The project separates GPU-heavy neural rendering on the server from interactive WebGL/WebCodecs composition on the client.

Problem

Client-only 3DGS rendering is constrained by browser memory, mobile GPU capacity, and the need to interact with local objects. A practical web system also needs to align server-rendered neural RGB-D output with client-side mesh geometry in the same view.

Approach

  • Server-side neural renderer outputs synchronized color and log-depth frames.
  • Transport service streams the encoded RGB-D interface to the browser.
  • WebCodecs and WebGL decode, render local mesh content, and perform screen-space depth fusion.
  • Coordinate-space alignment and depth normalization connect neural scenes with browser-side interaction.

Work

  • Hybrid client-server architecture for web neural graphics.
  • Color-and-depth streaming interface for reconstructed scenes.
  • Depth-based fusion between Gaussian splats and mesh geometry.
  • Browser-side collision and interaction experiments for neural rendering systems.

Results

  • Web3D 2025 Best Paper Award.
  • Compressed-depth path reduces payload from 4197 KB/frame to 219 KB/frame.
  • Throughput improves from 8.0 FPS to 35.2 FPS in the 1080p LivingLab evaluation.
  • Hybrid streaming keeps complex iPhone scenes runnable where client-only WebGL exceeds memory.

Publication

StreamSplat: A Hybrid Client-Server Architecture for Neural Graphics using Depth-based Fusion on the Web
The 30th ACM International Conference on 3D Web Technology (Web3D 2025), Siena, Italy.

Authors: Sehyeon Park, Yechan Yang, Myeongseong Kim, Byounghyun Yoo.

Recognition

The paper received the Best Paper Award at Web3D 2025.

Demo

Materials

Graphical abstract showing cross-platform web access, hybrid neural rendering, coordinate-space alignment, depth-based fusion, and depth-based collision

Final main figure showing server-side neural renderer, color/depth H.264 streaming, WebCodecs/WebGL client fusion, and WebGL mesh interaction

Animated RGB-D fusion preview with local RGB, server RGB, fusion RGB, and fusion depth views

Animated credit-card collision test with neural stream, local Open3D mesh, and fused view

Animated credit-card depth-fusion test showing local object movement and fused neural view

Animated close-range credit-card collision test showing depth-aware local mesh fusion

Animated ruler interaction test showing browser-side mesh movement and fused depth response

Depth normalization failure case showing incorrect server/client depth ordering before fusion correction

Depth-based collision demonstration and viewpoint failure heatmap from follow-up depth-fusion experiments

Prev
Speedometer Interface Development
Next
Guacamole — LLM Video Infringement Detection