CoDTS (AAAI 2025)

CoDTS: Enhancing Sparsely Supervised Collaborative Perception with a Dual Teacher-Student Framework (Paper)

Features:

• Dataset Support

  • OPV2V (ICRA 2022)
  • V2X-Sim 2.0 (NeurIPS 2021)
  • DAIR-V2X (CVPR 2022)
  • V2V4Real (CVPR 2023)

• Collaborative perception methods (Used in paper)

  • Attentive Fusion [ICRA 2022]
  • F-Cooper [SEC 2019]
  • DiscoNet [NeurIPS 2021]

• Visualization support

  • BEV visualization
  • 3D visualization

Installation

This code is based on CoAlign, so I recommend you visit CoAlign Installation Guide to learn how to install this repo.

Or you can refer to OpenCOOD data introduction and OpenCOOD installation guide to prepare data and install CoAlign. The installation is totally the same as OpenCOOD.

Dataset and Sparse Label

1. Download datasets

• DAIR-V2X Download

  Note: We use the Complemented Annotations for DAIR-V2X-C from Website

• V2V4Real Download

• OPV2V Download

• V2X-Sim 2.0 Download

2. Download sparse labels

The sparse label in this paper is in Google Drive, you can also produce sparse labels with codes in the "data_preparation" folder.

3. Put the folders of sparse labels for different datasets in the following:

CoDTS 
├── data_preparation # scripts to generate sparse labels
├── dataset # root of your dataset
│   ├── my_dair_v2x
│       ├── v2x_c
│         ├── cooperative-vehicle-infrastructure
│           ├── cooperative/label_world_sparse
│           ├── infrastructure-side/label/virtuallidar_sparse
│           ├── vehicle-side/label/lidar_sparse
│   ├── v2vreal
│       ├── train
│       ├── train_sparse # sparse label
│       ├── validate
│   ├── OPV2V
│       ├── test
│       ├── train
│       ├── train_sparse # sparse label
│       ├── validate
│   ├── V2X-Sim-2.0 # original v2xsim dataset
│   ├── v2xsim2_info # the label index file for v2xsim
│       ├── v2xsim_infos_test.pkl
│       ├── v2xsim_infos_train_plus.pkl # sparse label
│       ├── v2xsim_infos_train.pkl 
│       ├── v2xsim_infos_val.pkl
├── opencood # the core codebase

Training and Inference

Run the bash command in the terminal and modify the "DATASET" and "METHOD" parameters in the bash file to switch the dataset and collaborative detector.

Inference detectors

bash 0_test.bash

Train detectors with full/sparse labels only

bash 1_train_from_scratch.bash

Train detectors with sparse and pseudo labels

SSC3OD: Sparsely Supervised Collaborative 3D Object Detection from LiDAR Point Clouds (SMC 2023)

bash 2_ssl_pretrain.bash
bash 3_train_static_teacher.bash
bash 4_train_SSC3OD.bash

HINTED: Hard Instance Enhanced Detector with Mixed-Density Feature Fusion for Sparsely-Supervised 3D Object Detection (CVPR 2024)

bash 5_train_HINTED.bash

CoDTS: Enhancing Sparsely Supervised Collaborative Perception with a Dual Teacher-Student Framework (AAAI 2025)

bash 6_train_CoDTS.bash

Note: Both HINTED and CoDTS are based on static teachers in SSC3OD.

CoDTS generates pseudo labels based on the static teacher’s predictions, and the student is initialized with pre-trained encoder.

HINTED initially loads the static teacher’s parameters to the dynamic teacher and student before training.

The pre-trained encoder "DATASET_point_pillar_lidar_single_mae_0.7" and static teacher "DATASET_point_pillar_lidar_METHOD_ws_single_mae_0.7" in SSC3OD are in Google Driver.

Citation

@article{han2025codts,
  title={CoDTS: Enhancing Sparsely Supervised Collaborative Perception with a Dual Teacher-Student Framework},
  author={Yushan Han and Hui Zhang and Honglei Zhang and Jing Wang and Yidong Li},
  journal={arXiv preprint arXiv:2412.08344},
  year={2025}
}

Acknowledgment

This project is impossible without the code of OpenCOOD and CoAlign!