Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection

ICLR 2026 | Accepted

Overview

Fore-Mamba3D is a novel backbone architecture for 3D object detection that leverages Mamba-based linear modeling with foreground-enhanced encoding. Existing Mamba-based methods encode the entire scene including abundant background information, which is inefficient. Our method achieves superior performance by focusing on foreground voxels with enhanced contextual representation.

Key Contributions

1. Foreground-Enhanced Linear Encoding: Selects top-k foreground voxels via predicted scores, significantly reducing computational cost while maintaining accuracy

2. Regional-to-Global Sliding Window (RGSW): Aggregates information from local patches to the entire sequence, addressing response attenuation in sparse foreground voxels

3. SASFMamba Module: Enhances contextual representation through:

   • Semantic-Assisted Fusion (SAF): Enables long-range interactions between semantically similar voxels
   • State Spatial Fusion (SSF): Recovers geometric information lost in linearization

Method Overview

The backbone consists of four stages, each containing an instance selection block and a downsampling block. The instance selection block includes foreground voxel sampling, RGSW strategy, and SASFMamba encoder for effective foreground-focused 3D object detection.

Installation

git clone https://github.com/yourusername/Fore-Mamba3D.git
cd Fore-Mamba3D

# Create conda environment
conda create -n fore-mamba3d python=3.10
conda activate fore-mamba3d

# Install dependencies
pip install -r requirements.txt

# Install torch with CUDA support (example for CUDA 11.8)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Dataset Preparation

KITTI Dataset

# Download KITTI dataset
# Place it in the following structure:
data/kitti/
├── training/
│   ├── calib/
│   ├── image_2/
│   ├── label_2/
│   └── velodyne/
└── testing/
    ├── calib/
    ├── image_2/
    └── velodyne/

nuScenes Dataset

# Download nuScenes dataset
# Place it in the following structure:
data/nuscenes/
├── samples/
├── sweeps/
├── maps/
└── v1.0-trainval/

Training & Evaluation

# Multi-GPU training
python -m torch.distributed.launch \
    --nproc_per_node=4 \
    train.py --cfg configs/kitti_train.yaml

Results

KITTI Validation Set

Method	Backbone	Car (Hard/Mod/Easy)	Pedestrian (Hard/Mod/Easy)	Cyclist (Hard/Mod/Easy)
PointPillars	MLP	79.1 / 75.0 / 68.3	52.1 / 43.5 / 41.5	75.8 / 59.1 / 52.9
LION	Mamba	88.6 / 78.3 / 77.2	67.2 / 60.2 / 55.6	83.0 / 68.6 / 63.9
VoxelMamba	Mamba	89.1 / 80.8 / 78.1	66.0 / 59.7 / 53.7	84.2 / 69.1 / 64.8
Fore-Mamba3D	Mamba	90.3 / 82.2 / 79.5	67.8 / 62.2 / 57.0	86.4 / 69.5 / 66.3

nuScenes Validation Set

Method	Backbone	mAP	NDS
HEDNet	Transformer	66.7	71.4
SAFDNet	Hybrid	66.3	71.0
Voxel-Mamba	Mamba	67.5	71.9
LION	Mamba	68.0	72.1
FSHNet	Hybrid	68.1	71.7
Fore-Mamba3D	Mamba	68.4	72.3

Waymo Open Dataset (L2)

Method	Backbone	Vehicle	Pedestrian	Cyclist	mAP
Voxel-RCNN	SpCNN	76.1 / 68.2	78.2 / 69.3	70.8 / 68.3	68.6
LION	Mamba	- / 67.0	- / 75.4	- / 71.9	71.4
Fore-Mamba3D	Mamba	76.3 / 67.8	82.1 / 75.6	72.8 / 72.2	71.9

Performance Highlights

• KITTI: Achieves state-of-the-art performance with +1.2% improvement over VoxelMamba on Car class
• nuScenes: Validation mAP of 68.4% and Test mAP of 70.1%, outperforming LION and other SOTA methods
• Waymo: Competitive results with 71.9% mAP, showing strong generalization across datasets

Citation

If you find this project useful in your research, please cite:

@misc{ning2026foremamba3dmambabasedforegroundenhancedencoding,
   title={Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection}, 
   author={Zhiwei Ning and Xuanang Gao and Jiaxi Cao and Runze Yang and Huiying Xu and Xinzhong Zhu and Jie Yang and Wei Liu},
   year={2026},
   eprint={2602.19536},
   archivePrefix={arXiv},
   primaryClass={cs.CV},
   url={https://arxiv.org/abs/2602.19536}, 
}

Acknowledgments

This work is partially supported by:

• National Natural Science Foundation of China (Grant No. 62376153, 62402318, 24Z990200676, 62376252)
• Zhejiang Province Leading Geese Plan (2025C02025, 2025C01056)
• Zhejiang Province Province-Land Synergy Program (2025SDXT004-3)

License

This project is released under the MIT License.

Contributing

We welcome contributions from the community. Please feel free to submit pull requests or open issues.