WaveComm

March 7, 2026 · View on GitHub

[ICRA2026] WaveComm: Lightweight Communication for Collaborative Perception via Wavelet Feature Distillation

This is the official implementation of WaveComm, a wavelet-based communication framework for collaborative perception.

Overview

WaveComm is a novel collaborative perception framework that achieves communication efficiency by exploiting the frequency structure of feature maps. The core idea is to decompose BEV feature maps using Discrete Wavelet Transform (DWT), transmit only compact low-frequency components, and reconstruct the full features at the receiver using a lightweight generator.

Video

Key Features

Wavelet Feature Distillation: Uses DWT to decompose features into low-frequency and high-frequency components. Only low-frequency components are transmitted to reduce bandwidth.
Multi-Scale Distillation (MSD) Loss: A hybrid loss function combining:
- Reconstruction Loss (pixel-level)
- SSIM Loss (structural-level)
- Perceptual Loss (semantic-level)
- Adversarial Loss (distributional-level)
Lightweight Generator: A lightweight neural network that reconstructs high-frequency details from transmitted low-frequency components.
Communication efficiency: Reduces communication volume to 86.3% (OPV2V) and 87.0% (DAIR-V2X) of the original while maintaining state-of-the-art performance.

Method

Wavelet Feature Distillation Module

The Wavelet Feature Distillation Module consists of two parts:

Feature Decomposition: Using DWT to decompose BEV features into low-frequency and high-frequency components. Low-frequency components retain most semantic information and global structure.
Feature Reconstruction: A lightweight generator reconstructs the full feature map from the transmitted low-frequency components. The generator is optimized using MSD Loss to ensure high-fidelity reconstruction.

Performance

OPV2V Dataset

Method	Camera-based			LiDAR-based
Method	AP50	AP70	Comm	AP50	AP70	Comm
No Collaboration	0.405	0.216	0.0	0.782	0.634	0.0
F-Cooper	0.469	0.219	22.0	0.763	0.481	24.0
DiscoNet	0.517	0.234	22.0	0.882	0.737	24.0
AttFusion	0.529	0.252	22.0	0.878	0.751	24.0
V2X-ViT	0.603	0.289	22.0	0.917	0.790	24.0
CoBEVT	0.571	0.261	22.0	0.935	0.821	24.0
HM-ViT	0.643	0.370	22.0	0.950	0.873	24.0
WaveComm (ours)	0.681	0.451	19.0	0.965	0.926	21.0

DAIR-V2X Dataset

Method	Camera-based			LiDAR-based
Method	AP30	AP50	Comm	AP30	AP50	Comm
No Collaboration	0.014	0.004	0.0	0.421	0.405	0.0
F-Cooper	0.115	0.026	23.0	0.723	0.620	23.0
DiscoNet	0.083	0.017	23.0	0.746	0.685	23.0
AttFusion	0.094	0.021	23.0	0.738	0.673	23.0
V2X-ViT	0.198	0.057	23.0	0.785	0.521	23.0
CoBEVT	0.182	0.042	23.0	0.787	0.692	23.0
HM-ViT	0.163	0.044	23.0	0.818	0.761	23.0
WaveComm (ours)	0.274	0.123	20.0	0.831	0.790	20.0

Installation

Prerequisites

Python 3.8+
PyTorch 1.12.0+
CUDA 11.6+

Step 1: Create Environment

conda create -n wavecomm python=3.8
conda activate wavecomm
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.6 -c pytorch -c conda-forge

Step 2: Install Dependencies

pip install -r requirements.txt

Step 3: Install Spconv

pip install spconv-cu116  # match your cudatoolkit version

Step 4: Compile Bbx IoU CUDA Version

python opencood/utils/setup.py build_ext --inplace

Step 5: Install This Project

python setup.py develop

Data Preparation

OPV2V Dataset

Please refer to OpenCOOD. You also need to download additional-001.zip for camera modality.

DAIR-V2X Dataset

Download from DAIR-V2X. Use complemented annotation.

Create a dataset folder and organize as follows:

WaveComm/
└── dataset/
    ├── my_dair_v2x/
    │   ├── v2x_c/
    │   ├── v2x_i/
    │   └── v2x_v/
    └── OPV2V/
        ├── additional/
        ├── test/
        ├── train/
        └── validate/

Training

Train the Model

python opencood/tools/train.py -y ${CONFIG_FILE} [--model_dir ${CHECKPOINT_FOLDER}]

Arguments:

-y or --hypes_yaml: Path to training configuration file
--model_dir (optional): Path to checkpoint for fine-tuning or continuing training

Train with DDP

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --use_env opencood/tools/train_ddp.py -y ${CONFIG_FILE}

Testing

python opencood/tools/inference.py --model_dir ${CHECKPOINT_FOLDER} [--fusion_method intermediate]

Project Structure

WaveComm/
├── opencood/
│   ├── hypes_yaml/          # Configuration files
│   ├── models/              # Model architectures
│   │   └── comm_modules/    # Communication modules
│   ├── data_utils/          # Data loading and preprocessing
│   ├── fuse_modules/        # Feature fusion modules
│   ├── loss/                # Loss functions
│   ├── utils/               # Utility functions
│   └── tools/               # Training and testing scripts
├── dataset/                 # Dataset directory
├── assets/                  # assets for README
└── README.md

Citation

@article{wavecomm2026,
  title={WaveComm: Lightweight Communication of BEV Feature Maps for Collaborative Perception via Wavelet Feature Distillation},
  author={},
  journal={},
  year={2026}
}

Acknowledgments

This project is built upon the OpenCOOD and HEAL frameworks. We thank the authors for their excellent work.