Compact HD Map Construction via Douglas-Peucker Point Transformer

Authors: Ruixin Liu, Zejian Yuan

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on the public nu Scenes dataset demonstrate that our method overwhelms current SOTAs. Extensive ablation studies validate each component of our methods.
Researcher Affiliation Academia Ruixin Liu, Zejian Yuan* Institute of Artificial Intelligence and Robotics, Xi an Jiaotong University, China sweetylrx@stu.xjtu.edu.cn, yuan.ze.jian@xjtu.edu.cn
Pseudocode Yes Algorithm 1: DP point generation.
Open Source Code Yes Codes will be released at https://github.com/sweety121/DPFormer.
Open Datasets Yes Experiments are conducted on the large-scale nu Scenes (Caesar et al. 2020) dataset
Dataset Splits Yes we follow (Li et al. 2022a; Liu et al. 2023b; Liao et al. 2023) to train on the 700 scenes (28130 samples) from the training set and test on the 150 scenes (6019 samples) from the validation set.
Hardware Specification Yes All experiments are conducted on a single Nvidia RTX 3090.
Software Dependencies No The paper mentions software components like ResNet50 and AdamW optimizer, but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We train DPFormer with a batch size of 4 (each containing K = 6 images). The Adam W (Loshchilov and Hutter 2019) optimizer is adopted with a learning rate of 1.25e 4 for single-card training. According to data statistics shown in Figure 2, we set the maximum number of DP points as N = 8, the maximum number of all points as M = 20, and the maximum number of map elements as J = 50. Besides, the loss coefficients αc1, αc2, αp and αs are set to 2, 0.5, 5, and 0.5, respectively. Ablation studies are trained with 50 epochs.