TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning
Authors: Dongming Wu, Jiahao Chang, Fan Jia, Yingfei Liu, Tiancai Wang, Jianbing Shen
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To answer this question, we conduct detailed ablation studies on detection performance by varying the backbones. It shows that the topology performances are constantly improved with stronger detection. All experiments are conducted on the popular driving topology reasoning benchmark, Open Lane-V2, showing Topo MLP reaches state-of-the-art performance. |
| Researcher Affiliation | Collaboration | 1 Beijing Institute of Technology, 2 University of Science and Technology of China, 3 MEGVII Technology, 4 SKL-IOTSC, University of Macau |
| Pseudocode | No | No pseudocode or clearly labeled algorithm block was found in the paper. |
| Open Source Code | Yes | Code is at https://github.com/wudongming97/Topo MLP. |
| Open Datasets | Yes | The experiments are conducted on the Open Lane-V2 (Wang et al., 2023). Open Lane-V2 is a large-scale perception and reasoning dataset for scene structure in autonomous driving. |
| Dataset Splits | No | The paper states the total size of the subsets (1000 scenes each) but does not explicitly provide the specific training, validation, and test dataset splits needed for reproduction (e.g., percentages or sample counts for each split). |
| Hardware Specification | Yes | All the experiments are trained for 24 epochs on 8 Tesla A100 GPUs with a batch size of 8 if not specified. |
| Software Dependencies | No | The paper mentions using YOLOv8, but does not provide specific version numbers for YOLOv8 or other key software components like deep learning frameworks (e.g., PyTorch, TensorFlow) or their versions. |
| Experiment Setup | Yes | All images are resized into the same resolution of 1550 2048, and are downsampled with a ratio of 0.5. [...] The lane query number is set to NL = 300, and the number of control points is 4. [...] For lane detection loss Ldetl, the weight of the classification part is 1.5, and the weight of the regression part is 0.2. [...] The learning rate is initialized with 2.0 10 4 and decayed with cosine annealing policy (Loshchilov & Hutter, 2016). We adopt the HSV augmentation and grid mask strategy for training. All the experiments are trained for 24 epochs on 8 Tesla A100 GPUs with a batch size of 8 if not specified. |