Sketch and Refine: Towards Fast and Accurate Lane Detection
Authors: Chao Chen, Jie Liu, Chang Zhou, Jie Tang, Gangshan Wu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our SRLane can run at a fast speed (i.e., 278 FPS) while yielding an F1 score of 78.9%. Extensive experiments on benchmarks demonstrate that our SRLane could obtain remarkable performance with breakneck speed. |
| Researcher Affiliation | Academia | Chao Chen, Jie Liu*, Chang Zhou, Jie Tang, Gangshan Wu State Key Laboratory for Novel Software Technology, Nanjing University, China {chenchao, zhouchang}@smail.nju.edu.cn, {liujie,tangjie,gswu}@nju.edu.cn |
| Pseudocode | No | The paper describes its methodology in text and through diagrams, but it does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at: https://github.com/passerer/SRLane. |
| Open Datasets | Yes | Experiments are conducted on three popular lane datasets in the literature: Tusimple (Tu Simple 2017), CULane (Pan et al. 2018), and Curvelanes (Xu et al. 2020). |
| Dataset Splits | Yes | CULane is one of the largest lane detection datasets, which comprises 88,880 frames, 9,675 frames, 34,680 frames for training, validation, and testing, respectively. |
| Hardware Specification | Yes | For a fair comparison, the inference latency is measured on our machine with an AMD EPYC 7232P CPU and an NVIDIA Titan Xp GPU. |
| Software Dependencies | No | In the current implementation, all operators in the model are based on Py Torch (Paszke et al. 2019). The paper mentions PyTorch but does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | The local direction map is resized to 4 10 regardless of the input resolution, which means the number of proposals L is fixed as 40. The groups of lane segments is set to 6 by default. Training loss encompasses l1 loss for direction estimation, cross entropy loss for attention weights, focal loss (Lin et al. 2017b) for lane classification, and iou loss (Zheng et al. 2022) for lane regression. The overall loss is given by: L = wcls Lcls + wreg Lreg + wdir Ldir + wattn Lattn, where the loss weights are set as wcls = 2.0, wreg = 1.0, wdir = 0.05, and wattn = 0.05. We use Adam W (Loshchilov and Hutter 2017) as the optimizer in conjunction with a cosine learning rate scheduler. The initial learning rate is set to 10 3 with 800 iterations of linear warm-up. The batch size is set to 40 and images are resized to 800 320. Data augmentation for training includes random flipping, affine transformation, color jittering and JPEG compression. |