PolarMix: A General Data Augmentation Technique for LiDAR Point Clouds

Authors: Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, Ling Shao

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that Polar Mix achieves superior performance consistently across different perception tasks and scenarios.
Researcher Affiliation Collaboration 1School of Computer Science and Engineering, Nanyang Technological University 2Mohamed bin Zayed University of Artificial Intelligence 3Terminus Group, China
Pseudocode Yes Algorithm 1 Polar Mix.
Open Source Code Yes Code is available at https://github.com/xiaoaoran/polarmix
Open Datasets Yes The first is Semantic KITTI [1]... The second is nu Scenes-lidarseg [12] dataset... The third is Semantic POSS [30]... Syn Li DAR [46] is a synthetic Li DAR point cloud dataset...
Dataset Splits Yes Semantic KITTI: We follow the widely-adopted split and use sequences 00-07, 09-10 as the training set and sequence 08 for validation. ... nu Scenes-lidarseg: We follow the officially split of training data and validation data. ... Semantic POSS: We follow the official benchmark setting, i.e. sequence 03 for validation and the rest for training.
Hardware Specification Yes We conducted experiments with a single Tesla 2080Ti GPU for Mink Net and SPVCNN and a Tesla V100 GPU for Rand LA-Net and Cylinder3D.
Software Dependencies No The paper mentions using open-source repositories for various networks (e.g., Mink Net, SPVCNN, Rand LA-Net, Cylinder3D, Open PCDet) but does not provide specific version numbers for any software dependencies like libraries or frameworks.
Experiment Setup Yes We adopt the default training hyper-parameters in the open-source repositories for all four networks, and the only modification is the batch size for SPVCNN and Mink Net (we change it to 8). ... For augmentation with scene-level swapping, we randomly crop 180 sectors from 360 for [α, β] for point swapping. ... We set δ1, δ2 as 0.5, 1, respectively.