DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors

Authors: Xiaze Zhang, Ziheng Ding, Qi Jing, Yuejie Zhang, Wenchao Ding, Rui Feng

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to evaluate our proposed Deep Point Map and current state-of-the-art algorithms on benchmark datasets.
Researcher Affiliation Academia 1School of Computer Science, Shanghai Key Laboratory of Intelligent Information Processing, Fudan University 2Academy for Engineering and Technology, Fudan University 3Shanghai Collaborative Innovation Center of Intelligent Visual Computing
Pseudocode No The paper describes procedures but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes 1Source code: https://github.com/Zhang Xiaze/Deep Point Map
Open Datasets Yes Semantic KITTI (Behley et al. 2019, 2021) is a widely used benchmark dataset... KITTI-360 (Liao, Xie, and Geiger 2022) includes... Mul Ran (Kim et al. 2020) is a range dataset... KITTI-Carla (Deschaud 2021) is a synthetic dataset...
Dataset Splits Yes We use the first 6 sequences as training-set. ... The split of training and testing set is based on the frame amount of approx. 6:4, without any manual picking.
Hardware Specification Yes on 6 RTX 3090 GPU.
Software Dependencies No The paper mentions Open3D library and Adam W optimizer, but specific version numbers for these software dependencies are not provided.
Experiment Setup Yes with Adam W optimizer (Reddi, Kale, and Kumar 2019), initial learning rate lr = 10 3, weight decay wd = 10 4, cosine lr scheduler, on 6 RTX 3090 GPU. For all tasks, the network is trained for 12 epochs. ... We use the average of loss in both directions, with the hyper-parameters of εpositive = 1m, εoffset = 2m, τ = 0.1, λ1, λ2, λ3 = (1.0, 0.1, 1.0).