RESOLVE: A Multi-Resolution and Multi-Modal Dataset for Roadside Cooperative Perception

[ECCV 2026] The official codebase for the paper "RESOLVE: A Multi-Resolution and Multi-Modal Dataset for Roadside Cooperative Perception"

📢 Announcements

Stay up to date with the latest news, updates, and important notices regarding RESOLVE:

2026/06/18: Paper accepted at ECCV 2026 🎉.
2026/06/20: The v0.1.0 dataset and benchmark code for 3D object detection are released. The currently supported detection models are as follows: PointPillars, SECOND, CenterPoint, TransFusion, VoxSeT, DSVT, Voxel Mamba, LION, BEVFusion, UniTR
2026/06/22: The benchmark code for 3D multi-object tracking is released. The currently supported tracking models are as follows: AB3DMOT, CenterPoint, SimpleTrack, Poly-MOT, MCTrack
2026/06/24: The benchmark code for agent-level cooperative perception is released. The currently supported cooperative models are as follows: AttFuse, F-Cooper, CoBEVT, V2X-ViT

✅ TODO

Release the RESOLVE dataset
Release the benchmark code for 3D object detection
Release the benchmark code for 3D multi-object tracking
Release the benchmark code for cooperative perception

📦 Data Download

The RESOLVE Dataset is available upon registration. Please complete the Data Request Form to receive the download link via email.

After downloading and decompressing the data, please organize the data to the following structure:

DATA_ROOT
├── data
│   ├── sunlakes
│   │   │   │── 📂 2026-01-04_18_02_01-export
│   │   │   │   │── 📂 rosbag2_2025_10_14-07_21_21
│   │   │   │   │   │── 📂 16_16b_sync
│   │   │   │   │   │    │── 📄 1760451628-250275702.bin # point cloud fused by two low resolution lidar
│   │   │   │   │   │    └   ...
│   │   │   │   │   │── 📂 64_64b_sync
│   │   │   │   │   │── 📂 128_128b_sync
│   │   │   │   │   │── 📂 ouster64b_sync
│   │   │   │   │   │    │── 📄 1760451628-250275702.bin # point cloud captured by the low-resolution LiDAR at the northwest corner of the intersection
│   │   │   │   │   │    └   ...
│   │   │   │   │   │── 📂 ouster64_sync
│   │   │   │   │   │── 📂 ouster128b_sync
│   │   │   │   │   │── 📂 ouster128_sync
│   │   │   │   │   │── 📂 rslidar16b_sync
│   │   │   │   │   │── 📂 rslidar16_sync
│   │   │   │   │   │── 📂 axis1_sync
│   │   │   │   │   │    │── 🖼️ 1760451681-076657049.jpg # image captured by the camera located south of the intersection.
│   │   │   │   │   │    └   ...
│   │   │   │   │   │── 📂 axis2_sync
│   │   │   │   │   │── 📂 axis3_sync
│   │   │   │   │   │── 📂 axis4_sync
│   │   │   │   │   │── 📂 calibration_txt
│   │   │   │   │   │    │── 📂 axis1__ouster128b
│   │   │   │   │   │    │    │── 📊 1760451681-076657049__1760451628-150432431_calibration.txt
│   │   │   │   │   │    │    └   ...
│   │   │   │   │   │    │── 📂 axis2__ouster128b
│   │   │   │   │   │    │── 📂 axis3__ouster128b
│   │   │   │   │   │    │── 📂 axis4__ouster128b
│   │   │   │   │── 📊 HSQJN_3.json # annotation information file
│   │   │   │   └   ...
│   │   │   └   ...
├── pcdet
├── tools

🏋️ Getting Started

Requirements

All the codes are tested in the following environment:

Ubuntu 22.04
Python 3.10.19
PyTorch 2.8.0
CUDA 12.8
spconv v2.x

3D Object Detection

We implement 3d object detection based on OpenPCDet. Please follow OpenPCDet's official instructions to set up the environment first. Please note that for Voxel Mamba and LION, you need to install mamba-ssm separately by following the instruction.

Then generate the data infos by running the following command and data at different resolutions can be generated by setting the sensor_type parameter:

#  os128: high-resolution LiDAR data, os64: medium-resolution LiDAR data, rs16: low-resolution LiDAR data
python tools/create_sunlakes_data_v3.py --data_root ../data/sunlakes  --sensor_type os128

After generating the data infos, organize the data to follow the nuScenes directory structure:

DATA_ROOT
├── data
│   ├── sunlakes
│   │   │   │── 📂 2026-01-04_18_02_01-export
│   │   │   │   │── 📂 v1.0-trainval
│   │   │   │   │   │── 📂 sunlakes_gt_database
│   │   │   │   │   │    │── 📄 0_car_0.bin
│   │   │   │   │   │    └   ...
│   │   │   │   │   │── 📂 sunlakes_img_gt_database
│   │   │   │   │   │    │── 🖼️ 0_car_0.jpg
│   │   │   │   │   │    └   ...
│   │   │   │   │   │── 📊 sunlakes_infos_train.pkl
│   │   │   │   │   │── 📊 sunlakes_infos_val.pkl
│   │   │   │   │   │── 📊 sunlakes_dbinfos_train.pkl
│   │   │   │── 📂 lidar_pc_with_ts  # each point cloud segment is rotated to fit within a regular rectangular region.
│   │   │   │   │── 📂 rosbag2_2025_10_14-07_21_21
│   │   │   │   │   │── 📂 16_16b_sync  # the resolution of the generated LiDAR data is controlled by the `sensor_type` parameter.
│   │   │   │   │   │    │── 📄 1760451628-250275702.bin
│   │   │   │   │   │    └   ...
│   │   │   │   └   ...
├── pcdet
├── tools

When training a model at a specific resolution, ensure that DATA_PATH in cfgs/dataset_configs/sunlakes_dataset.yaml is set to the directory corresponding to that resolution.

# train with multiple GPUs
sh scripts/dist_train.sh 4 --cfg_file cfgs/sunlakes_models/bevfusion.yaml

# train with a single GPU:
python train.py --cfg_file ${CONFIG_FILE}

To test and evaluate the pretrained models:

# test with multiple GPUs
sh scripts/dist_test.sh 4 --cfg_file ${CONFIG_FILE} --batch_size ${BATCH_SIZE} --ckpt ${CKPT}

3D Multi-Object Tracking

As a prerequisite, the object detection algorithm should be executed to generate the corresponding detection results results_nusc.json. Also refer to the original repositories for instructions on setting up the environments required by each tracking algorithm.

After setup, replace the paths for data_root and pred_json with the local paths on your machine, and then run tracking script:

cd mot/MCTrack
# training and evaluation
python tracking.py --data_root your_path/sunlakes_infos_train.pkl --pred_json your_path/results_nusc.json

Cooperative Perception

We implement agent-level cooperative perception based on OpenCOOD. Please follow OpenCOOD's official instructions to set up the environment first.

Next, run the following commands to generate the data infos. By configuring the ego_sensor and cooperative_agents parameters, you can generate data for different agents and LiDAR resolutions.

cd cooperative/opencood/
#  os128/os128b: high-resolution LiDAR agents, os64/os64b: medium-resolution LiDAR agents, rs16/rs16b: low-resolution LiDAR agents
python tools/create_sunlakes_opencood.py --data_root ../data/sunlakes  --output_root your_output_path --ego_sensor os128b --cooperative_agents os128

After generating the data infos, organize the data to follow the directory structure:

DATA_ROOT
├── data
│   ├── sunlakes
│   │   │   │── 📂 your_output_path
│   │   │   │   │── 📂 train
│   │   │   │   │   │── 📂 2026-01-04_18_02_01-export__68fce85d185a9dc2c167cae6 
│   │   │   │   │   │    │── 📂 0 # ego_sensor data
│   │   │   │   │   │    │    │── 📄 1760718487-550083650.pcd  # point cloud from one agent at a specific timestamp
│   │   │   │   │   │    │    │── 📄 1760718487-550083650.yaml # label information from one agent at a specific timestamp
│   │   │   │   │   │    │    └   ...
│   │   │   │   │   │    │── 📂 1 # cooperative_agents data
│   │   │   │   │   └   ...
│   │   │   │   │── 📂 validate
│   │   │   │   │── 📂 openpcdet_eval # used to support OpenPCDet-style evaluation.
│   │   │   │   │   │── 📂 v1.0-trainval
│   │   │   │   │   │    │── 📊 sunlakes_infos_train.pkl
│   │   │   │   │   │    │── 📊 sunlakes_infos_val.pkl
├── pcdet
├── tools

We provide a script for training various cooperative perception methods on the RESOLVE dataset. Before training, update root_dir and validate_dir in the corresponding configuration files under cooperative/opencood/hypes_yaml with the local dataset paths on your machine.

For example, to train a model using 4 GPUs:

#  MODEL: no_fusion | late_fusion | early_fusion | attentive_fusion | fcooper | cobevt | v2xvit
sh run_sunlakes_4gpu.sh train ${MODEL}

🔍 Models Zoo

3D Object Detection Benchmarks

Model	Modality	LiDAR Backbone	Low Resolution	Mid Resolution	High Resolution
PointPillars	LiDAR	Sparse Convolution	75.1 / 71.2	79.7 / 72.6	80.6 / 73.6
SECOND	LiDAR	Sparse Convolution	68.2 / 66.5	76.6 / 71.2	78.0 / 73.1
CenterPoint	LiDAR	Sparse Convolution	79.9 / 73.9	86.4 / 79.9	87.4 / 80.9
TransFusion-L	LiDAR	Sparse Convolution	82.5 / 76.3	86.6 / 80.1	89.1 / 82.4
VoxSeT	LiDAR	Transformer	87.4 / 82.9	89.1 / 81.6	90.1 / 83.1
DSVT	LiDAR	Transformer	85.9 / 81.5	94.1 / 87.0	94.6 / 87.5
Voxel Mamba	LiDAR	Mamba	85.7 / 81.7	94.9 / 88.5	95.4 / 89.1
LION	LiDAR	Mamba	88.1 / 84.2	95.1 / 89.8	95.9 / 91.1
BEVFusion	LiDAR + Camera	Sparse Convolution	86.3 / 79.8	92.6 / 84.8	93.1 / 86.4
UniTR	LiDAR + Camera	Transformer	89.7 / 83.7	94.4 / 87.3	94.9 / 87.4

Agent-level Cooperative Perception Benchmarks

Model	Low Resolution	Mid Resolution	High Resolution
No Fusion	31.6	50.3	52.9
Early Fusion	48.6	61.0	66.0
Late Fusion	31.0	56.8	58.0
AttFuse	43.7	67.4	72.9
F-Cooper	36.8	63.2	68.9
CoBEVT	43.5	64.0	69.5
V2X-ViT	43.0	67.5	71.5

📝 License

Code: Licensed under the MIT License. See LICENSE file for details.
Dataset: Licensed under the Creative Commons Attribution 4.0 International CC BY-NC-ND 4.0. You must give appropriate credit; Cannot be used for commercial purposes; You may not distribute modified versions of the dataset.

📖 Citation

If you use RESOLVE Dataset, please cite:

@misc{resolve2026,
      title={RESOLVE: A Multi-Resolution and Multi-Modal Dataset for Roadside Cooperative Perception}, 
      author={Shaozu Ding and Linan Song and Marco De Vincenzi and Dajiang Suo},
      year={2026},
      eprint={2606.31895},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.31895}, 
}

Name		Name	Last commit message	Last commit date
Latest commit History 20 Commits
assets		assets
cooperative		cooperative
mot		mot
pcdet		pcdet
tools		tools
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
setup.py		setup.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

RESOLVE: A Multi-Resolution and Multi-Modal Dataset for Roadside Cooperative Perception

📢 Announcements

✅ TODO

📦 Data Download

🏋️ Getting Started

Requirements

3D Object Detection

3D Multi-Object Tracking

Cooperative Perception

🔍 Models Zoo

3D Object Detection Benchmarks

Agent-level Cooperative Perception Benchmarks

📝 License

📖 Citation

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

RESOLVE: A Multi-Resolution and Multi-Modal Dataset for Roadside Cooperative Perception

📢 Announcements

✅ TODO

📦 Data Download

🏋️ Getting Started

Requirements

3D Object Detection

3D Multi-Object Tracking

Cooperative Perception

🔍 Models Zoo

3D Object Detection Benchmarks

Agent-level Cooperative Perception Benchmarks

📝 License

📖 Citation

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages