(ICCV 2025) OmniSAM: Omnidirectional Segment Anything Model for UDA in Panoramic Semantic Segmentation

Ding Zhong^1,3,*, Xu Zheng^1,4,*, Chenfei Liao¹, Yuanhuiyi Lyu¹, Jialei Chen⁵, Shengyang Wu³, Linfeng Zhang⁶, Xuming Hu^1,2,†

¹AI Thrust, HKUST(GZ), ²CSE, HKUST, ³UMich, ⁴INSAIT, ⁵Nagoya University, ⁶SJTU

* Equal contribution. † Corresponding author.

---

📄 Abstract

Segment Anything Model 2 (SAM2) has emerged as a strong base model in various pinhole imaging segmentation tasks. However, when applying it to 360° domain, the significant field-of-view (FoV) gap between pinhole (70° × 70°) and panoramic images (180° × 360°) poses unique challenges. Two major concerns for this application include:

1. Inevitable distortion and object deformation brought by the large FoV disparity between domains.
2. The lack of pixel-level semantic understanding that the original SAM2 cannot provide.

To address these issues, we propose a novel OmniSAM framework, which makes the first attempt to apply SAM2 for panoramic semantic segmentation. Specifically, to bridge the first gap, OmniSAM first divides the panorama into sequences of patches. These patches are then treated as image sequences in similar manners as in video segmentation tasks. We then leverage the SAM2’s memory mechanism to extract cross-patch correspondences that address the cross-FoV dependencies, improving feature continuity and prediction consistency along mask boundaries. For the second gap, OmniSAM fine-tunes the pretrained image encoder and reinitializes the mask decoder for semantic prediction. An FoV-based prototypical adaptation module with dynamic pseudo label update mechanism is also introduced to enhance the alignment of memory and backbone features, thereby improving model generalization ability across different sizes of source models. Extensive experimental results demonstrate that OmniSAM outperforms the state-of-the-art methods by large margins, e.g., 79.06% (10.22↑) on SPin8-to-SPan8, 62.46% (6.58↑) on CS13-to-DP13.

🏗️ Framework

Preparation

📁 Data

Before getting started, please download the following datasets:

• Cityscapes
• SynPASS
• DensePASS
• Stanford2D3D

Then organize them into the following structure.

⚠️ Note: The data/ folder should be placed at the same level as the OmniSAM/ root directory.

data/
├── cityscapes
│   ├── leftImg8bit
|   |   ├── test 
|   |   ├── train
|   |   └── val
│   └── gtFine
|       ├── test 
|       ├── train
|       └── val
├── Stanford2D3D
│   ├── area_1
│   ├── area_2
│   ├── area_3
│   ├── area_4
│   ├── area_5a
│   ├── area_5b
│   └── area_6
├── SynPASS
│   ├── img
│   │   ├── cloud
│   │   ├── fog
│   │   ├── rain
│   │   └── sun
│   └── semantic
│       ├── cloud
│       ├── fog
│       ├── rain
│       └── sun
├── DensePASS
|   ├── gtFine
|   └── leftImg8bit
└── WildPASS
    └── leftImg8bit
OmniSAM/

🔧 Environment Setup

Create a new Conda environment and activate it:

conda create -n omnisam python=3.10
conda activate omnisam

Download the SAM2 weights and place them in the sam/checkpoints directory. For details, refer to the official SAM2 Installation Guidelines.

Install dependencies:

pip install -r requirements.txt

🚀 Getting Started

Please execute the following commend for source domain model training on multiple gpus:

torchrun --nnodes=1 --nproc_per_node=2 --master_port=29500 train.py --batch_size=8 --lr=1e-4 --num_epochs=10 --dataset='Cityscapes' --use_mem_bank --exp_name 'exp1' --backbone 'sam2_s' 

After you obtain a source model, you can use adaptation.py for domain adaptation:

python adaptation.py

To evaluate our model, please execute the following commend:

python validation.py --backbone='sam2_s' --dataset='DensePASS' --use_mem_bank

You can download our trained model from (https://drive.google.com/drive/folders/121HJ_8HXqTw6ne1iAjpzEMnb1RdRtpl-?usp=sharing).

🤝 Acknowledgement

Thanks to SAM2 team for their awesome work and Cityscapes/SynPASS/DensePASS/Stanford2D3D teams for their valuable datasets!

Reference

If you find this work useful in your research, please feel free to cite our paper :)

@misc{zhong2025omnisamomnidirectionalsegmentmodel,
      title={OmniSAM: Omnidirectional Segment Anything Model for UDA in Panoramic Semantic Segmentation}, 
      author={Ding Zhong and Xu Zheng and Chenfei Liao and Yuanhuiyi Lyu and Jialei Chen and Shengyang Wu and Linfeng Zhang and Xuming Hu},
      year={2025},
      eprint={2503.07098},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.07098}, 
}