Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Object Segmentation Techniques in Volumetric Rendering for 3D Scene Understanding

Peng, Ziyin LU (2025) MAMM15 20251
Department of Design Sciences
Abstract
Neural radiance fields enable photorealistic 3D scene representation but lack semantic understanding for object-level segmentation in complex indoor environments. This work addresses 3D object segmentation challenges in indoor scenes by integrating 3D Gaussian Splatting with 2D segmentation masks. I analyze existing segmentation methods across visual quality, accuracy, and efficiency dimensions, then follow a state-of-the-art approach that leverages multi-view consistency strategies to enhance segmentation stability across viewpoints. The method is validated on both public datasets and self-captured indoor scenes, demonstrating accurate segmentation while maintaining rendering quality and real-time performance. Results show significant... (More)
Neural radiance fields enable photorealistic 3D scene representation but lack semantic understanding for object-level segmentation in complex indoor environments. This work addresses 3D object segmentation challenges in indoor scenes by integrating 3D Gaussian Splatting with 2D segmentation masks. I analyze existing segmentation methods across visual quality, accuracy, and efficiency dimensions, then follow a state-of-the-art approach that leverages multi-view consistency strategies to enhance segmentation stability across viewpoints. The method is validated on both public datasets and self-captured indoor scenes, demonstrating accurate segmentation while maintaining rendering quality and real-time performance. Results show significant improvements in multi-view consistency and robust performance in challenging scenarios with occlusions. The approach shows strong potential for VR applications, enabling interactive scene editing and immersive content creation. (Less)
Popular Abstract (Swedish)
Neurala strålningsfält möjliggör fotorealistisk 3D-scenrepresentation men saknar semantisk förståelse för objektnivåsegmentering i komplexa inomhusmiljöer. Detta arbete adresserar utmaningar inom 3D-objektsegmentering i inomhusscener genom att integrera 3D Gaussian Splatting med 2D-segmenteringsmasker. Jag analyserar befintliga segmenteringsmetoder utifrån dimensionerna visuell kvalitet, noggrannhet och effektivitet, och följer sedan en toppmodern metod som utnyttjar multi-view-konsistensstrategier för att förbättra segmenteringsstabilitet över olika synvinklar. Metoden valideras på både offentliga dataset och egeninhämtade inomhusscener, vilket demonstrerar noggrann segmentering samtidigt som renderingskvalitet och realtidsprestanda... (More)
Neurala strålningsfält möjliggör fotorealistisk 3D-scenrepresentation men saknar semantisk förståelse för objektnivåsegmentering i komplexa inomhusmiljöer. Detta arbete adresserar utmaningar inom 3D-objektsegmentering i inomhusscener genom att integrera 3D Gaussian Splatting med 2D-segmenteringsmasker. Jag analyserar befintliga segmenteringsmetoder utifrån dimensionerna visuell kvalitet, noggrannhet och effektivitet, och följer sedan en toppmodern metod som utnyttjar multi-view-konsistensstrategier för att förbättra segmenteringsstabilitet över olika synvinklar. Metoden valideras på både offentliga dataset och egeninhämtade inomhusscener, vilket demonstrerar noggrann segmentering samtidigt som renderingskvalitet och realtidsprestanda bibehålls. Resultaten visar betydande förbättringar inom multi-view-konsistens och robust prestanda i utmanande scenarier med ocklusioner. Metoden uppvisar stark potential för VR-applikationer, vilket möjliggör interaktiv scenredigering och skapande av immersivt innehåll. (Less)
Please use this url to cite or link to this publication:
author
Peng, Ziyin LU
supervisor
organization
course
MAMM15 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Gaussian Splatting, Segmentation, Indoor Scene, Computer Vision, Segment Anything Model, NeRF
language
English
id
9199135
date added to LUP
2025-06-16 12:36:04
date last changed
2025-06-16 12:36:04
@misc{9199135,
  abstract     = {{Neural radiance fields enable photorealistic 3D scene representation but lack semantic understanding for object-level segmentation in complex indoor environments. This work addresses 3D object segmentation challenges in indoor scenes by integrating 3D Gaussian Splatting with 2D segmentation masks. I analyze existing segmentation methods across visual quality, accuracy, and efficiency dimensions, then follow a state-of-the-art approach that leverages multi-view consistency strategies to enhance segmentation stability across viewpoints. The method is validated on both public datasets and self-captured indoor scenes, demonstrating accurate segmentation while maintaining rendering quality and real-time performance. Results show significant improvements in multi-view consistency and robust performance in challenging scenarios with occlusions. The approach shows strong potential for VR applications, enabling interactive scene editing and immersive content creation.}},
  author       = {{Peng, Ziyin}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Object Segmentation Techniques in Volumetric Rendering for 3D Scene Understanding}},
  year         = {{2025}},
}