3D Harmonic Loss: Towards Task-consistent and Time-friendly 3D Object Detection on Edge for V2X Orchestration

Notification

Our proposed 3D Harmonic Loss can be applied to many lidar-based 3D object detection methods for solving inconsistency problem as shown below.

Paper in IEEE: PDF (IEEE T-VT)

Our implementation is relied on mmdetection3D

Environment

python = 3.7
pytorch = 1.6
CPU: i7-10700K
GPU: RTX-2080Ti
other requriements are same as in mmdetection3D

Dataset Preparation

Please follow the mmdetection3D to convert KITTI Dataset and Waymo Dataset

Training (MMdetection3D)

Note that our 3D harmonic loss optimization can be implemented to train almost all anchor-based 3D detectors without inference time cost. Please get familar with mmdetection3D-format config in advance, and then you can check configs/_base_/models/hv_second_secfpn_kitti_harmonic_loss.py as an example to customize other model configs, and use our revised anchor-head with harmonic loss for 3D detectors. You can follow official document of mmdetection3D to know how to configurate configs and train models.

Training (OpenPCDet)

will be released soon...

Test

You can follow official document of mmdetection3D to test models.

TODO Lists

• Readme Completion
• Paper Preprinted
• Support KITTI Dataset
• Support Waymo Dataset
• Support OpenPCDet benchmark