Hydro3DNet: Spatial-Attention Transformers for Hydrology-Aware Point Clouds

This repo is the official implementation of our paper Hydro3DNet: Spatial-Attention Transformers for Hydrology-Aware Point Clouds. Hydro3DNet is an end-to-end multimodal 3D object detection framework specifically designed for fixed waterway environments.

Highlights: -[25-09-06] Code of TROUT is released (SOTA).

Changelog

[2025-09-06] Hydro3DNet v0.1.0 is released.

[2025-12-30] Added support for the FSHNet baseline (configs and training scripts included).

[2026-03-25] Added support for the MAFF-Net baseline (configs and training scripts included).

[2026-03-29] Added support for the CAGroup3D baseline (configs and training scripts included).

Introduction

Hydro3DNet integrates hydrological information with LiDAR point clouds through an attention-based Transformer mechanism to enhance feature representation and improve detection performance.: 1)The proposed framework introduces a Hydro Augmented Voxel Feature Encoding (VFE) module, which embeds dynamic hydrological perception and boundary point augmentation strategies to enrich the spatial geometry features of waterway targets. 2)A Hydro Former Head module utilizes spatial attention and a Transformer point cloud encoder to integrate Bird's-Eye-View (BEV) features with region-of-interest (RoI) point clouds, enabling high-precision object detection in complex water scenes. 3)Extensive experiments on the publicly available TROUT dataset demonstrate that Hydro3DNet achieves state-of-the-art performance. On the NVIDIA RTX 4080 GPU, the Ship detection accuracy Ship|mAP3D of Hydro3DNet reaches 91.7%, and the inference speed is 37.9 Hz.

• Renderings of Different Models

🏆Main Results

TROUT Dataset

Test set

	Present at	Stages	Speed(HZ)	Training time	Ship mAP_3D (test)	Infra mAP_3D (test)	download
SECOND	Sensors’18	one	73.9	2.2	51.6	98.6	second_trout_64M
PointPillar	CVPR’19	one	40.5	12.1	61.6	99.4	pointpillar_trout_59M
PointRCNN	CVPR’19	one	8.8	15.2	11.9	98.1	pointrcnn_trout_48M
Part-A2	TPAMI’20	one	42.9	8.2	87.6	99.9	PartA2_trout_766M
PV-RCNN	CVPR’20	two	9.1	25.1	88.1	99.8	pv_rcnn_trout_158M
CenterPoint	CVPR’21	one	57.9	0.9	84.5	99.9	centerpoint_trout_94M
Voxel-RCNN	AAAI’21	two	48.3	8.3	85.9	99.8	voxel_rcnn_trout_89M
ST3D-SecondIOU	CVPR’21	two	66.0	6.1	69.9	99.7	second_iou_trout_144M
PV-RCNN++	IJCV’22	two	12.4	44.9	88.5	99.9	pv_rcnn_plusplus_trout_169M
PillarNet	ECCV’22	one	73.9	10.5	88.1	99.8	pillarnet_trout_169M
VoxelNext	CVPR’23	one	87.2	3.9	85.7	99.9	voxelnext_trout_89M
DSVT-Voxel	CVPR’23	one	6.8	62.4	54.7	95.5	dsvt_voxel_trout_99M
Lion-RWKV	NeurIPS’24	one	7.6	86.9	5.5	94.7	second_with_lion_rwkv_64dim_trout_104M
Voxel-Mamba	NeurIPS’24	one	13.6	22.9	57.8	96.2	voxel_mamba_trout_252M
ScatterFormer	ECCV’24	one	12.4	12.7	7.4	95.4	scatterformer_trout_147M
FSHNet	CVPR’25	one	16.6	16.5	29.2	25.7	fshnet_trout_132M
MAFF-Net	RL’25	two	20.9	45.7	27.9	71.3	MAFF-Net_trout_132M
CAGroup3D	CVPR’25	two	13.3	53.4	48.7	25.6	CAGroup3D_trout_954M
Hydro3DNet(our)	-	two	37.9	12.5	91.7	99.9	hydro3Dnet_trout_197M

Hydro3DNet achieves state-of-the-art performance. On the NVIDIA RTX 4080 GPU, the Ship detection accuracy Ship|mAP3D of Hydro3DNet reaches 91.7%, and the inference speed is 37.9 Hz. We hope that our Hydro3DNet can provide a potential group-free solution for efficiently handling sparse point clouds for 3D tasks.

🚀Usage

Installation

Please refer to INSTALL.md for installation.

Dataset Preparation

Please follow the instructions from GETTING_STARTED.md. We adopt the same data generation process.

Generate Hilbert Template

cd data
mkdir hilbert
python ./tools/hilbert_curves/create_hilbert_curve_template.py

You can also download Hilbert Template files from Google Drive or BaiduYun(code: kubf).

Training

In addition to Hydro3DNet, we are also integrating other models that emerged in the effect comparison of the papers (for example, DSVT, LION, Voxel-Mama...).

cd tool/
python train.py ${CONFIG_FILE} 

Give an example.

# 
cd tools
python train.py --cfg_file ./cfgs/trout_models/hydro3Dnet.yaml

# multi-gpu training
cd tools
bash scripts/dist_train.sh 8 --cfg_file ./cfgs/trout_models/hydro3Dnet.yaml 

/home/luxiaodong/Hydro3DNet-github/tools/cfgs/trout_models/hydro3Dnet.yaml

Test

# one-gpu testing
cd tools
python test.py --cfg_file ${CONFIG_FILE} --batch_size ${BATCH_SIZE} --ckpt ${CKPT}

Give an example.

python test.py  --cfg_file ./cfgs/trout_models/hydro3Dnet.yaml --batch_size 1 --ckpt ../output//cfgs/trout_models/hydro3Dnet/default/ckpt/latest_model.pth 

# multi-gpu testing
cd tools
bash scripts/dist_test.sh 8 --cfg_file <CONFIG_FILE> --ckpt <CHECKPOINT_FILE>

Citation

Please consider citing our work as follows if it is helpful.

@misc{trout2025,
    title={Hydro3DNet: Spatial-Attention Transformers for Hydrology-Aware Point Clouds},
    author={Xiaodong Lu, Sudong Xu},
    howpublished = {\url{https://github.com/serendipitylxd/Hydro3DNet}},
    year={2025}
}

Acknowledgments

Hydro3DNet is based on OpenPCDet.
We also thank the CenterPoint, TransFusion, OctFormer, Mamba, and … authors for their efforts.