MGMapNet: Multi-Granularity Representation Learning for End-to-End Vectorized HD Map Construction

Authors: Jing Yang, Minyue Jiang, Sen Yang, Xiao Tan, Yingying Li, Errui Ding, Jingdong Wang, Hanli Wang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that the proposed MGMap Net achieves state-of-the-art performances, surpassing Map TRv2 by 5.3 m AP on the nu Scenes dataset and 4.4 m AP on the Argoverse2 dataset, respectively. 4 EXPERIMENTS
Researcher Affiliation Collaboration 1 College of Electronic and Information Engineering, Tongji University, China 2 Baidu Inc.
Pseudocode No The paper describes the methodology using text and diagrams (Figure 2) but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about the release of source code or a link to a code repository.
Open Datasets Yes nu Scenes (Caesar et al., 2020) is a widely recognized dataset in the field of autonomous driving research, providing 1,000 scenes, each captured over a continuous 20-second interval. The Argoverse2 dataset (Wilson et al., 2023) contains multimodal data from 1,000 sequences, including high-resolution images from seven ring cameras and two stereo cameras, as well as Li DAR point clouds and map-aligned 6-Do F pose data.
Dataset Splits Yes For experimental purposes, the dataset is partitioned into 700 scenes comprising 28,130 samples for training, and 150 scenes containing 6,019 samples for validation. Results are reported on the validation set, with a focus on the same three map categories as identified in the nu Scenes dataset.
Hardware Specification Yes Our model is trained on 8 A100 GPUs with the batchsize of 2, utilizing the Adam W optimizer (Loshchilov et al., 2017) with the learning rate of 4 10 4.
Software Dependencies No The paper mentions the use of Adam W optimizer, ResNet50 as backbone, and LSS transformation, but does not provide specific version numbers for software libraries or dependencies like PyTorch, TensorFlow, or CUDA.
Experiment Setup Yes Our model is trained on 8 A100 GPUs with the batchsize of 2, utilizing the Adam W optimizer (Loshchilov et al., 2017) with the learning rate of 4 10 4. We adopt the Res Net50 (He et al., 2016) as the backbone and employ the LSS transformation (Philion & Fidler, 2020) with a single encoder layer for feature extraction. The one-to-many training strategy (Liao et al., 2023) is used, and the model is trained for 24 epochs on the nu Scenes dataset and 6 epochs on the Argoverse2 dataset. The hyperparameters are configured as Nq = 100, Nrep = 8, Np = 20, β1 = 5, β2 = 2, β3 = 0.005, β4 = 3, and β5 = 3.