SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar Object Detection
Authors: Ruoyu Xu, Zhiyu Xiang, Chenwei Zhang, Hanzhi Zhong, Xijun Zhao, Ruina Dang, Peng Xu, Tianyu Pu, Eryun Liu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results show that our SCKD outperforms the state-of-the-art methods, especially when large unlabeled data are available. ... Extensive experiments on the Vo D and ZJUODset datasets are carried out for evaluation. The results show that our radar-only student network is able to boost the performance of the baseline method by a large margin and outperforms the state-of-the-art methods. |
| Researcher Affiliation | Collaboration | 1Zhejiang University, China 2Zhejiang Provincial Key Laboratory of Multi-Modal Communication Networks and Intelligent Information Processing 3China North Artificial Intelligence & Innovation Research Institute EMAIL EMAIL |
| Pseudocode | No | The paper describes the proposed method using descriptive text and mathematical equations for loss functions, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/Ruoyu-Xu/SCKD |
| Open Datasets | Yes | We conduct experiments on the popular Vo D and ZJUODset datasets with accessible 4D radar and Lidar data. ... Vo D Dataset(Palffy et al. 2022) The Vo D dataset is currently the most popular 4D radar object detection dataset which includes Lidar, Radar, and Camera data. ... ZJUODset(Xu et al. 2023) The ZJUODset is a dataset for long-distance 3D object detection, with the farthest detection distance reaching up to 150 meters. |
| Dataset Splits | Yes | For the Vo D dataset, ... we divide the training and validation set into 5139 and 1296 frames, respectively. ... For the ZJUODset, ... we allocate 2660 frames for training and the subsequent 1140 frames for validation. We also use the rest 10640 unlabeled raw frames for semi-supervised distillation. |
| Hardware Specification | Yes | Two NVIDIA RTX 4090 GPUs are employed during the training and distillation, with the batch size set to 8. |
| Software Dependencies | No | The paper mentions implementing SCKD based on Open PCDet(Team 2020) and mmdetection3d(Contributors 2020) framework, but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | We employ Adam W optimizer for parameter update with an initial learning rate 0.001 and a weight decay factor 0.01. The learning rate is updated with a cyclical decay method, with maximum 0.01 and minimum 10 7. The retention threshold σ for the output distillation is set at 0.1, and the hyper-parameters α and β for the loss function are both set to 3 10 4. Two NVIDIA RTX 4090 GPUs are employed during the training and distillation, with the batch size set to 8. |