[ACM MM 2025] UIS-Mamba: Exploring Mamba for Underwater Instance Segmentation via Dynamic Tree Scan and Hidden State Weaken

Runmin Cong¹, Zongji Yu¹, Hao Fang^1†, Haoyan Sun¹, Sam Kwong^2
† Corresponding author
¹ School of Control Science and Engineering, Shandong University ² School of Data Science, Lingnan University

📖 Abstract

Underwater Instance Segmentation (UIS) is critical for underwater complex scene detection, but faces challenges like color distortion, blurred boundaries, and complex backgrounds. We propose UIS-Mamba—the first Mamba-based underwater instance segmentation model—equipped with two core modules: Dynamic Tree Scan (DTS) and Hidden State Weaken (HSW). UIS-Mamba achieves state-of-the-art (SOTA) performance on UIIS and USIS10K datasets while keeping parameters and computational complexity low.

📊 Experimental Results

Underwater Instance Segmentation (UIIS Dataset)

Method	Backbone	Params	mAP	AP₅₀	AP₇₅
WaterMask R-CNN	ResNet-50	54M	26.4	43.6	28.8
UIS-Mamba-T	UIS-Mamba-T	56M	29.4	46.7	31.3
WaterMask R-CNN	ResNet-101	67M	27.2	43.7	29.3
UIS-Mamba-S	UIS-Mamba-S	76M	30.4	48.6	33.2
USIS-SAM	ViT-H	700M	29.4	45.0	32.3
UIS-Mamba-B	UIS-Mamba-B	115M	31.2	49.1	34.5

Underwater Salient Instance Segmentation (USIS10K Dataset)

Method	Backbone	Params	Class-Agnostic mAP	Multi-Class mAP
WaterMask R-CNN	ResNet-50	54M	58.3	37.7
UIS-Mamba-T	UIS-Mamba-T	56M	62.2	42.1
WaterMask R-CNN	ResNet-101	67M	59.0	38.7
UIS-Mamba-S	UIS-Mamba-S	76M	63.1	44.5
USIS-SAM	ViT-H	701M	59.7	43.1
UIS-Mamba-B	UIS-Mamba-B	115M	63.8	46.2

🛠️ Environment Setup

Prerequisites

• Python 3.9+
• PyTorch 1.13.1+cu117 or higher
• MMDetection (for detection/segmentation heads)

Installation Steps

# Create conda environment
conda create -n uis-mamba python=3.9
conda activate uis-mamba

# Install PyTorch (CUDA 11.7)
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117

# Install other dependencies
pip install -r requirements.txt

# Install TreeScan Modules
cd third-party/TreeScan
pip install -v -e .

🚀 Train & Evaluate

1. Dataset Preparation & Pre-trained Weights

Download the two benchmark datasets and organize them as follows:

data/
├── UIIS/
│   ├── train/
│   │   ├── images/
│   │   └── annotations/
│   └── val/
│       ├── images/
│       └── annotations/
└── USIS10K/
    ├── train/
    ├── val/
    └── test/

• UIIS Dataset: Official Link
• USIS10K Dataset: Official Link

Pre-trained weights for UIS-Mamba variants (initialized with GrootV ImageNet-1K pre-trained weights) are available for download: Official Link

2. Training

Run training scripts for UIIS (instance segmentation) or USIS10K (salient instance segmentation):

# Train UIS-Mamba on UIIS/USIS10K (1 GPU)
python tools/train.py --config configs/vssm1/mask_rcnn_vssm_fpn_coco_tiny_ms_3x.py --work-dir you_dir_to_save_logs_and_models

3. Evaluation

Evaluate pre-trained models on validation/test sets:

# Evaluate on UIIS/USIS10K val set
python tools/test.py --config configs/vssm1/mask_rcnn_vssm_fpn_coco_tiny_ms_3x.py model_checkpoint_path --eval segm

📦 Model Zoo

Pre-trained weights for UIS-Mamba variants are available for download:

Model	Backbone	Dataset	mAP	Params	Download Link
UIS-Mamba-T	UIS-Mamba-T	UIIS	29.4	56M	ckpt
UIS-Mamba-S	UIS-Mamba-S	UIIS	30.4	76M	ckpt
UIS-Mamba-B	UIS-Mamba-B	UIIS	31.2	115M	ckpt
UIS-Mamba-T	UIS-Mamba-T	USIS10K	42.1	56M	ckpt
UIS-Mamba-S	UIS-Mamba-S	USIS10K	44.5	76M	ckpt
UIS-Mamba-B	UIS-Mamba-B	USIS10K	46.2	115M	ckpt

⭐ BibTeX

If you use UIS-Mamba in your research, please cite our paper:

@inproceedings{cong2025uis,
  title={UIS-Mamba: Exploring Mamba for Underwater Instance Segmentation via Dynamic Tree Scan and Hidden State Weaken},
  author={Cong, Runmin and Yu, Zongji and Fang, Hao and Sun, Haoyan and Kwong, Sam},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={343--352},
  year={2025}
}

❤️ Acknowledgement

Code is built upon MMDetection and GrootV.

☑️ LICENSE

The code is released under the MIT License.