CluB: Cluster Meets BEV for LiDAR-Based 3D Object Detection

Authors: Yingjie Wang, Jiajun Deng, Yuenan Hou, Yao Li, Yu Zhang, Jianmin Ji, Wanli Ouyang, Yanyong Zhang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted on Waymo and nu Scenes datasets, and our Clu B achieves state-of-the-art performance on both benchmarks. 4 Experiments In this section, we first make comparisons with the state-of-the-art methods on Waymo and nu Scenes datasets. Then, we conduct ablation studies to examine the effect of each component of the proposed Clu B.
Researcher Affiliation Academia Yingjie Wang1,2 , Jiajun Deng3 , Yuenan Hou2, Yao Li1, Yu Zhang1, Jianmin Ji1, Wanli Ouyang2, Yanyong Zhang1 1University of Science and Technology of China, 2Shanghai AI Laboratory, 3The University of Adelaide, AIML
Pseudocode No The paper describes its method in detail but does not include a clearly labeled pseudocode block or algorithm.
Open Source Code No The paper states 'Our code is built on MMdetection3D [6]' but does not provide a specific link or explicit statement about the public release of their own implementation code for Clu B.
Open Datasets Yes Extensive experiments are conducted on Waymo and nu Scenes datasets, and our Clu B achieves state-of-the-art performance on both benchmarks. Waymo Open Dataset [26] Nu Scenes [2] dataset
Dataset Splits Yes For Waymo Open Dataset [26] (WOD), 798, 202 and 150 sequences are used for training, validation and testing, respectively. Nu Scenes [2] dataset has 1,000 driving scenes, where 700, 150, and 150 scenes are chosen for training, validation and testing, respectively.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, memory, or specific cloud instance types used for running its experiments.
Software Dependencies No The paper mentions building code on 'MMdetection3D [6]' but does not provide specific version numbers for this or any other software dependencies.
Experiment Setup Yes For Waymo, we follow previous voxel-based methods [22, 35, 23, 21] to use point cloud range of [ 75.2m, 75.2m] [ 75.2m, 75.2m] [ 2.0m, 4.0m] with voxel size [0.1m, 0.1m, 0.15m] in x, y, and z-axes respectively. For nu Scenes, we use point cloud range of [ 51.2m, 51.2m] [ 51.2m, 51.2m] [ 5.0m, 3.0m] with voxel size [0.1m, 0.1m, 0.2m] in x, y, and z-axes respectively. We adopt the same data augmentation setting as [35], including random flipping, global scaling, global rotation, and groundtruth (GT) sampling [33] for Waymo dataset. ... We use the one-cycle [25] learning rate schedule and Adam W [16] optimizer with the maximal learning rate 0.001. In addition to the 12-epoch schedule for ablation studies, we adopt a longer schedule (36 epochs) to obtain the best performance on the validation and test set. During the evaluation, we use the NMS Io U threshold of [0.7, 0.25, 0.25].