EGFormer: Enhancing Group Attention for Off-Road Semantic Segmentation via Transition-Aware Refinement
Official implementation of "EGFormer: Enhancing Group Attention for Off-Road Semantic Segmentation via Transition-Aware Refinement".
EGFormer is an off-road semantic segmentation framework designed to improve boundary-aware perception in unstructured outdoor environments.
It introduces an edge-guided refinement strategy to better handle ambiguous transitions between terrain, vegetation, obstacles, and drivable regions.
- Edge-guided refinement for ambiguous off-road boundaries
- Transformer-based semantic segmentation architecture
- Robust to noisy labels and unclear class boundaries
- Support for RUGD and RELLIS-3D datasets
- MMSegmentation-based training and evaluation pipeline
- ROS support for real-world robotic perception
Overall architecture of EGFormer.
- RUGD
- RELLIS-3D
The dataset should be organized as follows:
EGFormer
├── data
│ ├── rellis
│ │ ├── test.txt
│ │ ├── train.txt
│ │ ├── val.txt
│ │ ├── annotation
│ │ │ ├── 00000
│ │ │ ├── 00001
│ │ │ ├── 00002
│ │ │ ├── 00003
│ │ │ └── 00004
│ │ └── image
│ │ ├── 00000
│ │ ├── 00001
│ │ ├── 00002
│ │ ├── 00003
│ │ └── 00004
│ │
│ └── rugd
│ ├── test_ours.txt
│ ├── test.txt
│ ├── train_ours.txt
│ ├── train.txt
│ ├── val_ours.txt
│ ├── val.txt
│ ├── RUGD_annotations
│ │ ├── creek
│ │ ├── park-1
│ │ ├── park-2
│ │ ├── park-8
│ │ ├── trail-1
│ │ ├── trail-3
│ │ ├── trail-7
│ │ ├── trail-9
│ │ ├── trail-15
│ │ └── village
│ └── RUGD_frames-with-annotations
│ ├── creek
│ ├── park-1
│ ├── park-2
│ ├── park-8
│ ├── trail-1
│ ├── trail-3
│ ├── trail-7
│ ├── trail-9
│ ├── trail-15
│ └── village
│
├── configs
├── tools
└── ...
