CrossGaussian
CrossGaussian
September 2024 ~ October 2025
🧑💼 Team Leader
📅 September 2024 ~ October 2025
🧑💼 Team Leader
🌐 ACM UIST 2025 Adjunct
🌐 ACM UIST 2025 Adjunct
📺️ IEEE ISMAR 2025 Research Demonstration
📺️ IEEE ISMAR 2025 Research Demonstration
Remote Collaboration System
Neural Rendering
Distributed System Architecture
User Study Design
CrossGaussian: Enhancing Remote Collaboration through
3D Gaussian Splatting and Real-time 360◦ Streaming
CrossGaussian: Enhancing Remote Collaboration through 3D Gaussian Splatting and Real-time 360◦ Streaming
Language
Abstract
Abstract
In remote collaboration systems, remote users often experience information asymmetry and limited interactivity when collaborating with on-site users using virtually reconstructed scenes of the physical environment. Real-time 360° camera streaming mitigates the narrow field-of-view limitations of conventional video conferencing systems by providing a wide-angle, fast-rendered view; however, the lack of depth information still restricts active and free spatial exploration. Conversely, offline CAD-based scene reconstruction allows free navigation but requires substantial time and cost to produce. To address these issues, this study introduces 3D Gaussian Splatting (3DGS) — a learning-based neural rendering technique capable of rapidly and accurately reconstructing large-scale physical environments with high responsiveness. CrossGaussian integrates real-time 360° video streaming and 3DGS-based large-scale scene reconstruction through an automated pipeline, thereby presenting the first room-scale remote collaboration design space that enables free-viewpoint exploration and novel visual interactions in remote collaborative environments.
In remote collaboration systems, remote users often experience information asymmetry and limited interactivity when collaborating with on-site users using virtually reconstructed scenes of the physical environment. Real-time 360° camera streaming mitigates the narrow field-of-view limitations of conventional video conferencing systems by providing a wide-angle, fast-rendered view; however, the lack of depth information still restricts active and free spatial exploration. Conversely, offline CAD-based scene reconstruction allows free navigation but requires substantial time and cost to produce. To address these issues, this study introduces 3D Gaussian Splatting (3DGS) — a learning-based neural rendering technique capable of rapidly and accurately reconstructing large-scale physical environments with high responsiveness. CrossGaussian integrates real-time 360° video streaming and 3DGS-based large-scale scene reconstruction through an automated pipeline, thereby presenting the first room-scale remote collaboration design space that enables free-viewpoint exploration and novel visual interactions in remote collaborative environments.
Introduction
Introduction
CrossGaussian is a first-author project I led from topic ideation to publication during my junior winter to senior fall in an HCI research lab. I defined the research direction, conducted literature reviews, designed the system, ran user studies, and presented the work at international conferences. In the early phase, I analyzed 20+ top-tier papers from CHI, UIST, and CVPR focused on remote collaboration systems, 3D reconstruction, and AI-based novel view synthesis. This review revealed the limitations of photogrammetry, NeRF, and Instant-NGP in remote collaboration, particularly high computational cost, slow processing, and limited interactivity. Based on this, I identified 3D Gaussian Splatting (3DGS) as a promising approach due to its explicit rendering and real-time performance. I independently designed an end-to-end prototype pipeline and developed it with co-authors. After building the prototype, I conducted a user study with 24 participants, collecting data via NASA-TLX, SUS, and custom questionnaires. Insights from early feedback led me to refine the research focus toward defining and exploring a design space for 3DGS-driven visualization and interaction techniques in remote collaboration. The work culminated in acceptance to the ACM UIST 2025 Poster Session and the ISMAR 2025 Demo Session, where I demonstrated the system live for three days. CrossGaussian demonstrates my ability to bridge cutting-edge AI rendering technology with human-centered design, delivering a functioning system that advances remote collaboration through HCI-driven insights.
CrossGaussian is a first-author project I led from topic ideation to publication during my junior winter to senior fall in an HCI research lab. I defined the research direction, conducted literature reviews, designed the system, ran user studies, and presented the work at international conferences. In the early phase, I analyzed 20+ top-tier papers from CHI, UIST, and CVPR focused on remote collaboration systems, 3D reconstruction, and AI-based novel view synthesis. This review revealed the limitations of photogrammetry, NeRF, and Instant-NGP in remote collaboration, particularly high computational cost, slow processing, and limited interactivity. Based on this, I identified 3D Gaussian Splatting (3DGS) as a promising approach due to its explicit rendering and real-time performance. I independently designed an end-to-end prototype pipeline and developed it with co-authors. After building the prototype, I conducted a user study with 24 participants, collecting data via NASA-TLX, SUS, and custom questionnaires. Insights from early feedback led me to refine the research focus toward defining and exploring a design space for 3DGS-driven visualization and interaction techniques in remote collaboration. The work culminated in acceptance to the ACM UIST 2025 Poster Session and the ISMAR 2025 Demo Session, where I demonstrated the system live for three days. CrossGaussian demonstrates my ability to bridge cutting-edge AI rendering technology with human-centered design, delivering a functioning system that advances remote collaboration through HCI-driven insights.
BACKGROUND
BACKGROUND
In co-located collaboration, participants can freely move, explore, and interact within the shared physical space. In remote collaboration, however, this autonomy is significantly constrained. Remote users depend on the on-site collaborator to look behind objects or change viewpoints within a video feed, increasing communication burden, causing unnecessary coordination, and ultimately limiting interaction. Some prior work mounts cameras on robotic platforms to provide spatial context, but despite this advantage, such systems often induce simulator sickness for remote users. Therefore, enabling free exploration of the physical environment for remote collaborators remains an open challenge.
In co-located collaboration, participants can freely move, explore, and interact within the shared physical space. In remote collaboration, however, this autonomy is significantly constrained. Remote users depend on the on-site collaborator to look behind objects or change viewpoints within a video feed, increasing communication burden, causing unnecessary coordination, and ultimately limiting interaction. Some prior work mounts cameras on robotic platforms to provide spatial context, but despite this advantage, such systems often induce simulator sickness for remote users. Therefore, enabling free exploration of the physical environment for remote collaborators remains an open challenge.
RESEARCH
RESEARCH
Real-time 360° video streaming partially mitigates viewpoint limitations by providing a wide field of view; however, the lack of depth information prevents users from actively understanding spatial structure or estimating object distance. While manually created 3D models offer another alternative, fully modeling every space is inefficient and costly. To address this, recent remote collaboration research has explored progressive reconstruction using camera-based photogrammetry. Yet, because this approach relies on image-based Structure-from-Motion (SfM) to produce surface-centric meshes, it suffers from limitations in resolution, accuracy, and responsiveness. More recently, Neural Radiance Fields (NeRF) have been adopted for remote collaboration, but their high computational cost still makes them unsuitable for real-time interactive environments.
Real-time 360° video streaming partially mitigates viewpoint limitations by providing a wide field of view; however, the lack of depth information prevents users from actively understanding spatial structure or estimating object distance. While manually created 3D models offer another alternative, fully modeling every space is inefficient and costly. To address this, recent remote collaboration research has explored progressive reconstruction using camera-based photogrammetry. Yet, because this approach relies on image-based Structure-from-Motion (SfM) to produce surface-centric meshes, it suffers from limitations in resolution, accuracy, and responsiveness. More recently, Neural Radiance Fields (NeRF) have been adopted for remote collaboration, but their high computational cost still makes them unsuitable for real-time interactive environments.
In contrast, the recently emerging camera-synthesized view–based neural rendering approach, 3D Gaussian Splatting (3DGS), represents a scene as a collection of numerous Gaussian primitives (defined by position, color, covariance, etc.) to enable fast rendering. Unlike NeRF, which encodes a scene implicitly through neural networks, 3DGS adopts an explicit and computation-efficient structure optimized for rapid processing. Owing to these characteristics, 3DGS offers significantly faster training speed, higher rendering performance, and superior scalability to large-scale or dynamic environments compared to NeRF. Building on this, our study integrates real-time contextual 360° video streaming with the fast, precise, and responsive capabilities of 3DGS for remote collaboration. Furthermore, we explore a room-scale design space for remote environment exploration and interactive techniques based on this integration.
In contrast, the recently emerging camera-synthesized view–based neural rendering approach, 3D Gaussian Splatting (3DGS), represents a scene as a collection of numerous Gaussian primitives (defined by position, color, covariance, etc.) to enable fast rendering. Unlike NeRF, which encodes a scene implicitly through neural networks, 3DGS adopts an explicit and computation-efficient structure optimized for rapid processing. Owing to these characteristics, 3DGS offers significantly faster training speed, higher rendering performance, and superior scalability to large-scale or dynamic environments compared to NeRF. Building on this, our study integrates real-time contextual 360° video streaming with the fast, precise, and responsive capabilities of 3DGS for remote collaboration. Furthermore, we explore a room-scale design space for remote environment exploration and interactive techniques based on this integration.
SYSTEM ARCHITECURE
SYSTEM ARCHITECURE
Following the research ideation, I led the end-to-end design and implementation of a distributed system architecture for prototype development. I integrated the entire data pipeline from real-time video capture to remote GPU training and rendering, implementing resource orchestration in a distributed computing environment through a Python-based automation framework with SSH/SFTP protocols. I designed a TCP-based streaming protocol and optimized GPU-accelerated encoding/decoding pipelines to improve network transmission stability and latency, while implementing a Unity Compute Shader-based rendering engine that leveraged 3DGS's explicit scene representation for novel visualization and interaction techniques including overlapping blending, depth analysis, and see-through exploration. Based on this foundation, I aimed to establish an environment for quantitatively validating system performance through user study design and analysis of standard HCI metrics such as NASA-TLX and SUS.
Following the research ideation, I led the end-to-end design and implementation of a distributed system architecture for prototype development. I integrated the entire data pipeline from real-time video capture to remote GPU training and rendering, implementing resource orchestration in a distributed computing environment through a Python-based automation framework with SSH/SFTP protocols. I designed a TCP-based streaming protocol and optimized GPU-accelerated encoding/decoding pipelines to improve network transmission stability and latency, while implementing a Unity Compute Shader-based rendering engine that leveraged 3DGS's explicit scene representation for novel visualization and interaction techniques including overlapping blending, depth analysis, and see-through exploration. Based on this foundation, I aimed to establish an environment for quantitatively validating system performance through user study design and analysis of standard HCI metrics such as NASA-TLX and SUS.
To integrate 3DGS and real-time camera streaming into a remote collaboration environment, I independently designed the full system pipeline.
1. Data Collection — Real-Time Synchronized Input: The system begins with real-time 360° video capture. Footage from an Insta360 camera is automatically stored locally and immediately referenced by a custom SDK plugin that forwards frames to a remote GPU server for multi-view optimization and 3D scene training. As soon as video is recorded, it is automatically passed into the image-based point-cloud generation pipeline, triggering reconstruction. Simultaneously, 360° video is streamed in real time via H.264 packets.
2. Reconstruction — Automated and Remote Gaussian Pipeline: To eliminate manual execution and path configuration during collaboration, I built an automated end-to-end pipeline connecting Unity and Python. A Python script (Putty.py) uses Paramiko to create SSH/SFTP connections, sending commands to the remote GPU server for training. Only video capture occurs locally; all heavy computation (SfM, Gaussian optimization, conversion) runs automatically on the remote server. Once training completes, result files are automatically retrieved and rendered in the collaboration environment.
3. Streaming — Reliable and Synchronized Transmission: Real-time video transmission is handled via a TCP-based streaming protocol. Each encoded frame is segmented into small chunks, and decoding occurs only after all chunks are received. ACK signals ensure reliable delivery with no duplication or loss, enabling low-latency, glitch-free 360° streaming synchronized with 3DGS output.
4. Rendering — Unified Visualization and Immersive Overlay: Finally, 360° video and Gaussian scenes are fused into a single space. Incoming video is decoded in real time using FFmpeg’s GPU decoder (h264_cuvid) and converted from NV12 to RGBA via NVIDIA NPP. A shader maps each fisheye pixel into spherical space, projecting the camera’s field of view onto a virtual dome for wide-FOV real-time streaming. Gaussian rendering results are composited on the same render target, producing a unified immersive scene.
To integrate 3DGS and real-time camera streaming into a remote collaboration environment, I independently designed the full system pipeline.
1. Data Collection — Real-Time Synchronized Input: The system begins with real-time 360° video capture. Footage from an Insta360 camera is automatically stored locally and immediately referenced by a custom SDK plugin that forwards frames to a remote GPU server for multi-view optimization and 3D scene training. As soon as video is recorded, it is automatically passed into the image-based point-cloud generation pipeline, triggering reconstruction. Simultaneously, 360° video is streamed in real time via H.264 packets.
2. Reconstruction — Automated and Remote Gaussian Pipeline: To eliminate manual execution and path configuration during collaboration, I built an automated end-to-end pipeline connecting Unity and Python. A Python script (Putty.py) uses Paramiko to create SSH/SFTP connections, sending commands to the remote GPU server for training. Only video capture occurs locally; all heavy computation (SfM, Gaussian optimization, conversion) runs automatically on the remote server. Once training completes, result files are automatically retrieved and rendered in the collaboration environment.
3. Streaming — Reliable and Synchronized Transmission: Real-time video transmission is handled via a TCP-based streaming protocol. Each encoded frame is segmented into small chunks, and decoding occurs only after all chunks are received. ACK signals ensure reliable delivery with no duplication or loss, enabling low-latency, glitch-free 360° streaming synchronized with 3DGS output.
4. Rendering — Unified Visualization and Immersive Overlay: Finally, 360° video and Gaussian scenes are fused into a single space. Incoming video is decoded in real time using FFmpeg’s GPU decoder (h264_cuvid) and converted from NV12 to RGBA via NVIDIA NPP. A shader maps each fisheye pixel into spherical space, projecting the camera’s field of view onto a virtual dome for wide-FOV real-time streaming. Gaussian rendering results are composited on the same render target, producing a unified immersive scene.
To integrate 3DGS and real-time camera streaming into a remote collaboration environment, I independently designed the full system pipeline.
1. Data Collection — Real-Time Synchronized Input: The system begins with real-time 360° video capture. Footage from an Insta360 camera is automatically stored locally and immediately referenced by a custom SDK plugin that forwards frames to a remote GPU server for multi-view optimization and 3D scene training. As soon as video is recorded, it is automatically passed into the image-based point-cloud generation pipeline, triggering reconstruction. Simultaneously, 360° video is streamed in real time via H.264 packets.
2. Reconstruction — Automated and Remote Gaussian Pipeline: To eliminate manual execution and path configuration during collaboration, I built an automated end-to-end pipeline connecting Unity and Python. A Python script (Putty.py) uses Paramiko to create SSH/SFTP connections, sending commands to the remote GPU server for training. Only video capture occurs locally; all heavy computation (SfM, Gaussian optimization, conversion) runs automatically on the remote server. Once training completes, result files are automatically retrieved and rendered in the collaboration environment.
3. Streaming — Reliable and Synchronized Transmission: Real-time video transmission is handled via a TCP-based streaming protocol. Each encoded frame is segmented into small chunks, and decoding occurs only after all chunks are received. ACK signals ensure reliable delivery with no duplication or loss, enabling low-latency, glitch-free 360° streaming synchronized with 3DGS output.
4. Rendering — Unified Visualization and Immersive Overlay: Finally, 360° video and Gaussian scenes are fused into a single space. Incoming video is decoded in real time using FFmpeg’s GPU decoder (h264_cuvid) and converted from NV12 to RGBA via NVIDIA NPP. A shader maps each fisheye pixel into spherical space, projecting the camera’s field of view onto a virtual dome for wide-FOV real-time streaming. Gaussian rendering results are composited on the same render target, producing a unified immersive scene.
DESIGN IMPLEMENTATION
DESIGN IMPLEMENTATION
Leveraging the explicit scene representation structure of 3DGS and its precise depth rendering capabilities at room scale, we explored a design space to enhance explorability and interactivity in remote collaboration environments. Inspired by existing Cross-Reality scene blending research, we designed the following key features for remote collaboration.
Leveraging the explicit scene representation structure of 3DGS and its precise depth rendering capabilities at room scale, we explored a design space to enhance explorability and interactivity in remote collaboration environments. Inspired by existing Cross-Reality scene blending research, we designed the following key features for remote collaboration.
Blending of Overlapping Scenes
Blending of Overlapping Scenes
Abrupt transitions between real-time streaming and 3DGS scenes can cause motion sickness and reduced presence. To mitigate this, our system implements a feature that visually separates yet blends 3DGS scenes with 360-degree video streaming. By adjusting the transparency of each overlaid scene and applying color scaling techniques, users can maintain real-time environmental context (360-degree stream) while simultaneously exploring with free viewpoints (3DGS). This reduces cognitive load during context switching and preserves presence. The overlay structure also enables visual distinction of non-salient regions through color scaling of 3DGS scenes or pixel value adjustments of 360-degree footage. Similar to Gruenefeld et al.'s adjustable scene blending, users can customize their optimal collaboration experience by controlling the blending ratio between 3DGS and 360-degree video to balance realism and exploration freedom.
Abrupt transitions between real-time streaming and 3DGS scenes can cause motion sickness and reduced presence. To mitigate this, our system implements a feature that visually separates yet blends 3DGS scenes with 360-degree video streaming. By adjusting the transparency of each overlaid scene and applying color scaling techniques, users can maintain real-time environmental context (360-degree stream) while simultaneously exploring with free viewpoints (3DGS). This reduces cognitive load during context switching and preserves presence. The overlay structure also enables visual distinction of non-salient regions through color scaling of 3DGS scenes or pixel value adjustments of 360-degree footage. Similar to Gruenefeld et al.'s adjustable scene blending, users can customize their optimal collaboration experience by controlling the blending ratio between 3DGS and 360-degree video to balance realism and exploration freedom.
Occlusion-Aware Exploration
Occlusion-Aware Exploration
Automatic Occlusion Detection and Visualization: Because the camera in a remote space captures the scene from a single viewpoint, there is a fundamental limitation in that it cannot visualize areas occluded by structures such as walls or pillars. For example, when an on-site worker needs to inspect equipment located behind a column, the 360° camera alone cannot reveal the hidden region. To address this, our system utilizes the 3D spatial information of the 3D Gaussian Splatting (3DGS) model to automatically detect and visualize occluded areas based on the camera’s position within the physical environment. Specifically, the system first computes which regions are blocked by surrounding structures relative to the current position and orientation of the 360° camera in the 3DGS model. It then compares the depth values of adjacent pixels to estimate each pixel’s precise depth and pseudo-normal direction, determining which parts correspond to shadows or occluded regions. Using Unity compute shaders and HLSL, the system performs real-time GPU-based shadow computation to quickly identify these occluded areas and visually highlight them for the user. Through this approach, remote users can intuitively perceive and utilize the camera’s viewpoint coverage in the on-site environment for more effective collaboration.
Automatic Occlusion Detection and Visualization: Because the camera in a remote space captures the scene from a single viewpoint, there is a fundamental limitation in that it cannot visualize areas occluded by structures such as walls or pillars. For example, when an on-site worker needs to inspect equipment located behind a column, the 360° camera alone cannot reveal the hidden region. To address this, our system utilizes the 3D spatial information of the 3D Gaussian Splatting (3DGS) model to automatically detect and visualize occluded areas based on the camera’s position within the physical environment. Specifically, the system first computes which regions are blocked by surrounding structures relative to the current position and orientation of the 360° camera in the 3DGS model. It then compares the depth values of adjacent pixels to estimate each pixel’s precise depth and pseudo-normal direction, determining which parts correspond to shadows or occluded regions. Using Unity compute shaders and HLSL, the system performs real-time GPU-based shadow computation to quickly identify these occluded areas and visually highlight them for the user. Through this approach, remote users can intuitively perceive and utilize the camera’s viewpoint coverage in the on-site environment for more effective collaboration.
See-Through Capability: Our system provides see-through capabilities for remote 3D environments by leveraging depth information inherent in the 3DGS model. While photogrammetry relies on mesh-based representations with fixed surfaces, making transparency control difficult, Gaussian Splatting uses 3D Gaussians with alpha values, enabling natural semi-transparent rendering through alpha blending at the rendering stage. This allows users to see through objects and directly inspect spaces beyond them without complex viewpoint manipulation, enabling intuitive exploration and novel interactions without information loss caused by occlusion.
See-Through Capability: Our system provides see-through capabilities for remote 3D environments by leveraging depth information inherent in the 3DGS model. While photogrammetry relies on mesh-based representations with fixed surfaces, making transparency control difficult, Gaussian Splatting uses 3D Gaussians with alpha values, enabling natural semi-transparent rendering through alpha blending at the rendering stage. This allows users to see through objects and directly inspect spaces beyond them without complex viewpoint manipulation, enabling intuitive exploration and novel interactions without information loss caused by occlusion.
PERCEPTUAL EVALUATION
PERCEPTUAL EVALUATION
In this study, we conducted a user study to investigate the impact of reconstruction delay on users' sense of object presence and perceived affordance. While state-of-the-art 3D reconstruction techniques often exceed the 33ms real-time processing threshold, previous research has rarely addressed how such delays affect user perception. Using a within-subjects experimental design with 18 participants, we implemented four randomized delay conditions (0.15s, 1s, 10s, 60s), asking participants to observe objects under each condition and rate their perceived affordance and confidence in object presence on a 7-point Likert scale. Analysis using non-parametric statistical methods—Friedman test and post-hoc Wilcoxon signed-rank tests with Bonferroni correction—revealed that reconstruction delay significantly reduced perceived affordance (χ²(3)=28.4, p<0.001), with ratings of 5.8±1.6 at 0.15s, 5.7±1.4 at 1s, 5.2±1.4 at 10s, and 4.3±1.7 at 60s. Statistically significant differences were observed particularly between 0.15s-60s (p<0.001) and 1s-60s (p<0.01) conditions. Confidence in object presence showed a similar pattern (χ²(3)=31.2, p<0.001), with ratings of 6.2±1.2 at 0.15s, 5.8±1.3 at 1s, 5.1±1.5 at 10s, and 4.3±1.8 at 60s. Statistical analysis was performed using Python's SciPy library, and condition-wise distributions and significance were visualized using Matplotlib and Seaborn. Qualitative analysis revealed that most participants began losing confidence after 10 seconds, and at 60-second delays, the object felt disconnected from reality, significantly reducing their willingness to interact.
In this study, we conducted a user study to investigate the impact of reconstruction delay on users' sense of object presence and perceived affordance. While state-of-the-art 3D reconstruction techniques often exceed the 33ms real-time processing threshold, previous research has rarely addressed how such delays affect user perception. Using a within-subjects experimental design with 18 participants, we implemented four randomized delay conditions (0.15s, 1s, 10s, 60s), asking participants to observe objects under each condition and rate their perceived affordance and confidence in object presence on a 7-point Likert scale. Analysis using non-parametric statistical methods—Friedman test and post-hoc Wilcoxon signed-rank tests with Bonferroni correction—revealed that reconstruction delay significantly reduced perceived affordance (χ²(3)=28.4, p<0.001), with ratings of 5.8±1.6 at 0.15s, 5.7±1.4 at 1s, 5.2±1.4 at 10s, and 4.3±1.7 at 60s. Statistically significant differences were observed particularly between 0.15s-60s (p<0.001) and 1s-60s (p<0.01) conditions. Confidence in object presence showed a similar pattern (χ²(3)=31.2, p<0.001), with ratings of 6.2±1.2 at 0.15s, 5.8±1.3 at 1s, 5.1±1.5 at 10s, and 4.3±1.8 at 60s. Statistical analysis was performed using Python's SciPy library, and condition-wise distributions and significance were visualized using Matplotlib and Seaborn. Qualitative analysis revealed that most participants began losing confidence after 10 seconds, and at 60-second delays, the object felt disconnected from reality, significantly reducing their willingness to interact.
OUTCOME
OUTCOME
I collaborated with my advisor to write the paper using Overleaf for real-time LaTeX-based document collaboration, directly created system architecture and user scenario figures through GPT image generation prompt engineering and Photoshop post-processing, and designed the academic poster in Figma to complete the conference presentation materials. This research was accepted as a first-author poster at ACM UIST (ACM Symposium on User Interface Software and Technology) 2025, the premier conference in user interface and interaction technology. It was also accepted for the demo session at IEEE ISMAR (International Symposium on Mixed and Augmented Reality) 2025, the world's leading conference in augmented and mixed reality, where I conducted live demonstrations for three days. At both conferences, I received significant interest and positive feedback from renowned HCI researchers worldwide and professionals from global companies regarding the system's real-time performance, practical utility for remote collaboration, and the innovative nature of the 3DGS-based approach.
I collaborated with my advisor to write the paper using Overleaf for real-time LaTeX-based document collaboration, directly created system architecture and user scenario figures through GPT image generation prompt engineering and Photoshop post-processing, and designed the academic poster in Figma to complete the conference presentation materials. This research was accepted as a first-author poster at ACM UIST (ACM Symposium on User Interface Software and Technology) 2025, the premier conference in user interface and interaction technology. It was also accepted for the demo session at IEEE ISMAR (International Symposium on Mixed and Augmented Reality) 2025, the world's leading conference in augmented and mixed reality, where I conducted live demonstrations for three days. At both conferences, I received significant interest and positive feedback from renowned HCI researchers worldwide and professionals from global companies regarding the system's real-time performance, practical utility for remote collaboration, and the innovative nature of the 3DGS-based approach.
MATERIALS
MATERIALS
This project has been published as an Adjunct Proceedings paper in the ACM Digital Library. To access the Full Paper, please click the image on the right to be redirected to the publication page. The paper is available for free under Open Access.
This project has been published as an Adjunct Proceedings paper in the ACM Digital Library. To access the Full Paper, please click the image on the right to be redirected to the publication page. The paper is available for free under Open Access.