5363-UR Moving Object Segmentation in Driving Scenes Using 4D Gaussian Splatting

Location

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Document Type

Event

Start Date

20-11-2026 2:00 PM

Description

Moving-object segmentation helps distinguish dynamic objects from static structure in driving-scene reconstruction. Framewise predictions can vary over time, while reconstruction errors can also arise in static regions. We present a track-level approach using 4D Gaussian Splatting to combine motion evidence across video frames. SAM2 provides instance tracks, while a Periodic Vibration Gaussian model supplies velocity and depth information. We combine aggregated velocity with geometric free-space evidence to identify moving tracks, then apply a low-residual filter to reduce false detections. The pipeline assigns one motion decision per track without requiring manual motion annotations. Using shared hyperparameters, the full method achieves mean F1 scores of 0.738 on Waymo and 0.842 on KITTI-MOT. Its masks also improve reconstruction quality in four Gaussian Splatting methods. Remaining challenges include distant movers, weak motion cues, and fragmented instance tracks. These results support track-level reasoning as a practical approach to moving-object segmentation.

This document is currently not available here.

Share

COinS
 
Nov 20th, 2:00 PM

5363-UR Moving Object Segmentation in Driving Scenes Using 4D Gaussian Splatting

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Moving-object segmentation helps distinguish dynamic objects from static structure in driving-scene reconstruction. Framewise predictions can vary over time, while reconstruction errors can also arise in static regions. We present a track-level approach using 4D Gaussian Splatting to combine motion evidence across video frames. SAM2 provides instance tracks, while a Periodic Vibration Gaussian model supplies velocity and depth information. We combine aggregated velocity with geometric free-space evidence to identify moving tracks, then apply a low-residual filter to reduce false detections. The pipeline assigns one motion decision per track without requiring manual motion annotations. Using shared hyperparameters, the full method achieves mean F1 scores of 0.738 on Waymo and 0.842 on KITTI-MOT. Its masks also improve reconstruction quality in four Gaussian Splatting methods. Remaining challenges include distant movers, weak motion cues, and fragmented instance tracks. These results support track-level reasoning as a practical approach to moving-object segmentation.