Xu Han, Yiduo Wang, Yiran Wang, Viorela Ila
Australian Centre for Robotics, School of Aerospace, Mechanical and Mechatronic Engineering, The University of Sydney
| DAVIS2017-M | |||
|---|---|---|---|
|
|
|
|
| KITTI | |||
|
|
||
|
|
||
Overview of the proposed DynOSeg framework. DynOSeg computes a residual map from depth, camera poses, and optical flow. Dedicated encoders process the depth and residual maps, while a frozen DINOv2 encoder extracts high-level semantic features from the RGB image. Geometry-Aware Spatial Fusion (GSF) and Global–Local Semantic-Aware Fusion (GLSF) integrate these multimodal features. A segmentation decoder then produces moving-object masks from the fused representation.
The reference hardware and software environment is listed below:
| Component | Specification |
|---|---|
| Operating system | Ubuntu 24.04.4 LTS (Noble Numbat) |
| GPU | NVIDIA GeForce RTX 4090 |
| NVIDIA driver | 580.159.03 |
| NVCC Compiler | V12.8.61 |
The code is comming soon!








