V2X-DGPE

January 22, 2025 · View on GitHub

Image Description

V2X-DGPE: Addressing Domain Gaps and Pose Errors for Robust Collaborative 3D Object Detection.

Installation

You can follow the CoAlign installation steps provided in the Feishu documents:

  • Chinese version: Link
  • English version: Link

Alternatively, refer to the OpenCOOD data introduction and OpenCOOD installation guide for data preparation and installation steps for CoAlign. The installation process is the same as OpenCOOD, with the exception of some additional dependencies required by CoAlign.

Dataset Preparation

DAIR-V2X

  1. Prepare the DAIR-V2X-C dataset and then prepare the complemented annotations.
  2. Then process the data: 
    cd ~/V2X-DGPE
python opencood/data_utils/datasets/basedataset/dairv2x_basedataset.py
  3. The folder structure should look like this: 
    cooperative-vehicle-infrastructure/
├── cooperative/
├── gt_database_fusion/
├── infrastructure-side/
├── vehilce-side/
├── dairv2x_dbinfos_fusion.pkl
├── train.json
└── val.json

Training

Train the Teacher Model

bash opencood/tools/scripts/dist_train.sh 4 opencood/hypes_yaml/dairv2x/lidar_only/pointpillar_early_gtsample_multiscale.yaml early

Train the Student Model

bash opencood/tools/scripts/train_w_kd.sh opencood/hypes_yaml/dairv2x/lidar_only/pointpillar_pdd_distillation.yaml opencood/logs/v2x_dgpe_student intermediate

Train the Student Model (Noise)

bash opencood/tools/scripts/train_w_kd.sh opencood/hypes_yaml/dairv2x/lidar_only/pointpillar_pdd_distillation.yaml opencood/logs/v2x_dgpe_student_noise intermediate

Testing

Run the following command to test the student model:

python opencood/tools/inference.py --model_dir opencood/logs/v2x_dgpe_student --fusion_method intermediate
python opencood/tools/inference_w_noise.py --model_dir opencood/logs/v2x_dgpe_student_noise --fusion_method intermediate

Checkpoints

The V2X-DGPE teacher , student, student-noise, v2x-dgpe/opencood/logs/lcfm240602(for history input) models, can be found in opencood/logs.

Acknowledgements

Thanks to V2X-VIT and DI-V2X for their contributions. The code is built on CoAlign.