DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection

Authors: Xiang Li, Junbo Yin, Wei Li, Chengzhong Xu, Ruigang Yang, Jianbing Shen

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on the challenging DAIR-V2X and V2XSet benchmark datasets demonstrate DI-V2X achieves remarkable performance, outperforming all the previous V2X models.
Researcher Affiliation Collaboration 1School of Computer Science, Beijing Institute of Technology 2Inceptio 3SKL-IOTSC, CIS, University of Macau
Pseudocode No The paper includes architectural diagrams and flowcharts but no explicit pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/Serenos/DI-V2X.
Open Datasets Yes We employ the challenging DAIR-V2X (Yu et al. 2022) for evaluating our model and other SOTA approaches. We also evaluate our method on another V2X calibration dataset V2XSet (Xu et al. 2022a)
Dataset Splits Yes V2XSet (Xu et al. 2022a), which contains 6694 training data and 1920 validation data generated by the simulator. Extensive experiments on the challenging DAIR-V2X and V2XSet benchmark datasets demonstrate DI-V2X achieves remarkable performance, outperforming all the previous V2X models. Table 1: Comparison with state-of-the-art methods on DAIR-V2X val dataset.
Hardware Specification Yes All student models are trained on 4 NVIDIA Tesla V100 GPUs with a batch size of 4 for 40 epochs.
Software Dependencies No The paper mentions using 'Point Pillars' as a detector but does not specify version numbers for any software dependencies or libraries (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes We set the point cloud range to [ 100, 100] [ 40, 40] [ 3.5, 1.5] meters defined in the vehicle coordinate system with the voxel size as [0.4, 0.4, 5] meters along XY Z axes. All student models are trained on 4 NVIDIA Tesla V100 GPUs with a batch size of 4 for 40 epochs. We set λkd as 1 and the thresholds τl, τh as 0.2 and 0.8.