A Hybrid Global-Local Perception Network for Lane Detection

Authors: Qing Chang, Yifei Tong

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that our proposed method improves detection accuracy in various challenging scenarios, outperforming the state-of-the-art lane detection methods.
Researcher Affiliation Academia School of Mechanical Engineering, Nanjing University of Science and Technology, China qingchang@njust.edu.cn, tyf51129@aliyun.com
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not explicitly provide access to its source code via a link or statement of availability.
Open Datasets Yes We conduct experiments on three widely used benchmarks: CULane (Pan et al. 2018), LLAMAS (Behrendt and Soussan 2019), and Tu Simple (Tu Simple 2020).
Dataset Splits No The paper specifies training and testing image counts for CULane (88k training, 34k testing) and Tu Simple (3,626 training, 2,782 testing). While it mentions a "valid set" for LLAMAS, it does not provide specific split sizes (e.g., counts or percentages) for a distinct validation set across all datasets to reproduce the data partitioning.
Hardware Specification Yes Training and testing are both performed on Pytorch with one Tesla-V100 GPU.
Software Dependencies No The paper mentions "Pytorch" as the framework used, but does not provide specific version numbers for Pytorch or any other software dependencies.
Experiment Setup Yes All input images are resized to 320 800 during the training and testing phases. In the training process, we use Adam W (Loshchilov and Hutter 2017) optimizer with an initial learning rate of 1e-3 and cosine decay learning rate strategy (Loshchilov and Hutter 2016) with power set to 0.9. For the lane prior head, we set the number of lane prior proposals Np = 192, proposals feature dimension Cp = 64, and the number of points of each lane prior N = 72. For the global extraction head, the resized Hg, Wg are 10, 25, respectively, the channel Cg = 192 and reference points K = 4. The dilation rate in MHA is set as d = 2. The training numbers of epochs for CULane, Tusimple, and LLAMAS are 15, 80, and 20.