Online Map Vectorization for Autonomous Driving: A Rasterization Perspective

Authors: Gongjie Zhang, Jiahao Lin, Shuang Wu, yilin song, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that incorporating rasterization into map vectorization greatly enhances performance with no extra computational cost during inference, leading to more accurate map perception and ultimately promoting safer autonomous driving.
Researcher Affiliation Collaboration Black Sesame Technologies Nanyang Technological University, Singapore
Pseudocode No The paper describes its methods and framework through text and diagrams, but it does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Project Page: https://github.com/ZhangGongjie/MapVR
Open Datasets Yes nu Scenes Map (basic) [2], which consists of two line-shaped map classes (lane divider and road boundary) and one polygon-shaped map class (pedestrian crossing). This dataset setup aligns with prior works on map vectorization [16, 6, 26, 20]. 2. nu Scenes Map (extended) [2], an extension of nu Scenes Map (basic) that incorporates more complex map elements, such as intersection, stopline area, and carpark area. 3. Argoverse2 [49], a large-scale dataset featuring the same classes as nu Scenes Map (basic).
Dataset Splits Yes nu Scenes Map (basic) validation set. nu Scenes Map (extended) validation set. Argoverse2 validation set.
Hardware Specification Yes All experiments, unless otherwise stated, are conducted with 8x NVIDIA RTX 3090 GPUs. Results were obtained with 8x NVIDIA A100 GPUs under the same training setups.
Software Dependencies No The paper does not specify software dependencies with version numbers, such as programming language versions or library versions (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes All experiments, unless otherwise stated, are conducted with 8x NVIDIA RTX 3090 GPUs. As the proposed Map VR is a generic framework with no reliance on specific model architecture, we adopt Map TR [20], the state-of-the-art model for map vectorization, as the base model. Our implementation aligns with Map TR [20]. The spatial size of the HD map is 480 × 240, such that each pixel represents 0.125m. We dilate the rasterized polylines by 2 pixels on each side. The softness τ for line rasterization is studied with values 0.5, 1.0, 2.0, 4.0, 6.0. Models are trained for 24 or 110 epochs using ResNet-50 as backbone.