Learning Transferable Features for Point Cloud Detection via 3D Contrastive Co-training

Authors: Zeng Yihan, Chunwei Wang, Yunbo Wang, Hang Xu, Chaoqiang Ye, Zhen Yang, Chao Ma

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our proposed 3D-Co Co effectively closes the domain gap and outperforms the state-of-the-art methods by large margins. We construct new domain adaptation benchmarks using three large-scale 3D datasets. Experimental results show that our proposed 3D-Co Co effectively closes the domain gap and outperforms the state-of-the-art methods by large margins.
Researcher Affiliation Collaboration 1 Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Huawei Noah s Ark Lab {zengyihan,weiwei0224,yunbow,chaoma}@sjtu.edu.cn {xu.hang,yechaoqiang,yang.zhen}@huawei.com
Pseudocode Yes Algorithm 1: The learning procedure of 3D contrastive co-training (3D-Co Co)
Open Source Code No The paper does not include an unambiguous statement where the authors explicitly state they are releasing the code for the work described in this paper, nor does it provide a direct link to a source-code repository.
Open Datasets Yes We evaluate 3D-Co Co on three widely used Li DAR-based datasets, including Waymo [26], nu Scenes [1], and KITTI [6].
Dataset Splits Yes We set the maximum number of training epochs to 30 for KITTI and 20 for Waymo and nu Scenes, with a warm-up process taking half of the total epochs.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory used for running its experiments.
Software Dependencies No The paper mentions software components like "Voxel Net", "Point Pillars", and "Adam optimizer", but it does not specify version numbers for any of these or other key software libraries/frameworks (e.g., Python, PyTorch, TensorFlow, CUDA).
Experiment Setup Yes We set the voxel size to (0.1, 0.1, 0.15)m for Voxel Net and (0.1, 0.1)m for Point Pillars. We use the Adam optimizer [13] with a learning rate of 1.5 × 10−3. We set the maximum number of training epochs to 30 for KITTI and 20 for Waymo and nu Scenes, with a warm-up process taking half of the total epochs. For pseudo-labels generation, we apply a high-pass threshold of 0.7 to Io U to obtain foreground samples, and a low-pass threshold of 0.2 for background samples.