Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
SafeMap: Robust HD Map Construction from Incomplete Observations
Authors: Xiaoshuai Hao, Lingdong Kong, Rong Yin, Pengwei Wang, Jing Zhang, Yunfeng Diao, Shu Zhao
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that Safe Map significantly outperforms previous methods in both complete and incomplete scenarios, highlighting its superior performance and resilience. |
| Researcher Affiliation | Collaboration | 1Beijing Academy of Artificial Intelligence 2National University of Singapore 3Institute of Information Engineering, Chinese Academy of Sciences 4School of Computer Science, Wuhan University 5School of Computer Science, Hefei University of Technology 6Intelligent Interconnected Systems Laboratory of Anhui Province (Hefei University of Technology) 7Independent Researcher. Correspondence to: Yunfeng Diao <EMAIL>. |
| Pseudocode | No | The paper describes the methodology in prose and mathematical equations (e.g., Eq. 1, 2, 3, 4, 5) and includes architectural diagrams (Figure 2, 3), but it does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions that baseline models were retrained using "official model configurations provided by the open-source codebases" for fairness, but it does not state that the code for Safe Map itself is released or provide any link to its source code. |
| Open Datasets | Yes | The nu Scenes dataset (Caesar et al., 2020) contains 1,000 sequences of recordings collected by autonomous driving cars. The Argoverse2 dataset (Wilson et al., 2021) consists of 1,000 logs, each capturing 15 seconds of 20Hz RGB images from 7 cameras, 10Hz Li DAR sweeps, and a 3D vectorized map. |
| Dataset Splits | Yes | The Argoverse2 dataset (Wilson et al., 2021) consists of 1,000 logs, each capturing 15 seconds of 20Hz RGB images from 7 cameras, 10Hz Li DAR sweeps, and a 3D vectorized map. The dataset is split into 700 logs for training, 150 logs for validation, and 150 logs for testing. |
| Hardware Specification | Yes | Our proposed Safe Map framework is trained using four NVIDIA RTX 3090 GPUs. |
| Software Dependencies | No | All experiments utilize the Adam W optimizer (Loshchilov & Hutter, 2019) with a learning rate of 4.2 10 4, fine-tuning for 8 epochs, while hyperparameters λ1 and λ2 are set to 0.05 and 5, respectively. The paper mentions an optimizer by name but does not provide specific version numbers for any software libraries, frameworks, or programming languages used. |
| Experiment Setup | Yes | All experiments utilize the Adam W optimizer (Loshchilov & Hutter, 2019) with a learning rate of 4.2 10 4, fine-tuning for 8 epochs, while hyperparameters λ1 and λ2 are set to 0.05 and 5, respectively. The original images have a resolution of 1, 600 900, resized by a factor of 0.5 during training. We limit the maximum number of map elements per frame to 100, with each containing 20 points. The size of each Bird s-Eye-View (BEV) grid is set to 0.75 meters, and the transformer decoder is configured with two layers. For the G-PVR module, the Safe Map model undergoes fine-tuning for 8 epochs on the nu Scenes dataset and 2 epochs on the Argoverse2 dataset, with batch sizes set to 4 and 6, respectively. We set σ to 3 for all configurations. |