A Siamese Transformer with Hierarchical Refinement for Lane Detection
Authors: Zinan Lv, Dong Han, Wenzhe Wang, Danny Z Chen
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on three benchmark datasets of lane detection demonstrate that our proposed new method achieves state-of-the-art results with high accuracy and efficiency. Specifically, our method achieves improved F1 scores on the Open Lane dataset, surpassing the current best-performing method by 5.0 points. |
| Researcher Affiliation | Academia | Zinan Lv1 Dong Han2 Wenzhe Wang2 Danny Z. Chen3 1Shanghai Jiao Tong University 2Zhejiang University 3University of Notre Dame |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We will release our code and checkpoint of this paper, and all the datasets we use are open access. |
| Open Datasets | Yes | To demonstrate the effectiveness of our proposed method in realistic road environments, we evaluate it on three benchmark datasets: Open Lane [5], CULane [26], and Tusimple[34]. |
| Dataset Splits | Yes | Open Lane is a real-world large-scale lane detection dataset, which contains 160K and 40K images as the training and validation sets, respectively. The validation set consists of six realistic road scenarios and annotates 14 lane categories (including white dotted line, double yellow solid, left/right curb, and so on). CULane is a widely-used large dataset for lane detection including eight hard-to-detect scenarios in urban areas and on highways, with 88K and 34K images as the training and validation sets, respectively. |
| Hardware Specification | Yes | All the experiments are conducted on a machine with a single NVIDIA RTX3090 GPU with 24GB memory. |
| Software Dependencies | No | The paper mentions optimizers (AdamW) and learning rate strategies but does not provide specific version numbers for software dependencies such as deep learning frameworks (e.g., PyTorch, TensorFlow) or programming languages (e.g., Python). |
| Experiment Setup | Yes | In the optimization process, we adopt Adam W [21] and the cosine decay learning rate strategy [20] with the initial learning rate set to 6e-4. A batch size of 32 and training epoch numbers of 10, 20, and 90 are used for Open Lane, CULane, and Tusimple, respectively. All the input images are reshaped into 800 320 pixels each for both the training and inference stages. For the number of lane anchors in an image, we set it to 150. |