pith. sign in

arxiv: 1907.09438 · v1 · pith:E6KZSPXKnew · submitted 2019-07-22 · 💻 cs.CV

Multi-Class Lane Semantic Segmentation using Efficient Convolutional Networks

Shao-Yuan Lo , Hsueh-Ming Hang , Sheng-Wei Chan , Jing-Jhih Lin This is my paper

Pith reviewed 2026-05-24 18:05 UTC · model grok-4.3

classification 💻 cs.CV
keywords multi-class lane segmentationsemantic segmentationconvolutional networksfeature size selectiondilated convolutionsreal-time inferencelane detectionautonomous driving
0
0 comments X

The pith

Feature Size Selection and Degressive Dilation Blocks improve accuracy for distinguishing lane types in road images while keeping or increasing inference speed.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper targets multi-class lane semantic segmentation, where a network must label different kinds of lanes such as solid, dashed, or colored markings. It starts from the observation that lanes occupy small, narrow regions in typical driving photos and introduces Feature Size Selection to match feature map dimensions to those thin objects plus a Degressive Dilation Block built from dilated convolutions whose rates decrease across layers. Experiments on the resulting network show measurable accuracy gains over a baseline segmentation model at the same or higher frames per second, including real-time rates on high-resolution inputs. A reader would care because reliable lane-type labels matter for autonomous driving systems that must interpret road rules from camera data.

Core claim

The authors establish that inserting Feature Size Selection to extract thin lane features at appropriate resolutions and a Degressive Dilation Block of dilated convolutions with progressively smaller dilation rates into a convolutional segmentation backbone produces obvious accuracy gains on multi-class lane labeling while matching or exceeding the baseline inference speed and supporting real-time operation on high-resolution images.

What carries the argument

Feature Size Selection (FSS), which chooses feature map sizes suited to narrow lane widths, and the Degressive Dilation Block (DD Block), which stacks dilated convolutions using decreasing dilation rates to retain fine spatial detail.

If this is right

  • The network produces more accurate labels for distinct lane classes without added compute cost.
  • Real-time processing remains feasible on high-resolution driving imagery.
  • The two blocks can be dropped into existing segmentation architectures to raise multi-class output quality.
  • Lane feature extraction becomes more robust to the narrow geometry typical of road markings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same size-selection and degressive-dilation pattern might transfer to segmentation of other thin linear structures such as cracks or wires.
  • Pairing the blocks with video-frame recurrence could test whether temporal consistency further lifts accuracy in moving scenes.
  • Evaluating the method across weather, lighting, or camera-angle shifts would check whether the reported gains hold outside the training distribution.

Load-bearing premise

The design assumes lanes consistently appear as small-size and narrow-width objects in road scene images, which is used to justify the chosen feature sizes and the degressive dilation schedule.

What would settle it

If side-by-side tests on a standard multi-class lane dataset show the proposed blocks yield no accuracy increase or slower inference than the unmodified baseline, the performance claim would be falsified.

Figures

Figures reproduced from arXiv: 1907.09438 by Hsueh-Ming Hang, Jing-Jhih Lin, Shao-Yuan Lo, Sheng-Wei Chan.

Figure 1
Figure 1. Figure 1: The proposed techniques applied to an existing semantic segmentation network for multi-class lane semantic segmentation. speed, but it does not distinguish different types of lanes. Zang et al. [21] used a CNN to find two-class lane pixels, yellow lines and white lines, yet their method can only process very small size, 32×32, images. Recognizing different types of lanes, such as yellow line, red line, dou… view at source ↗
Figure 2
Figure 2. Figure 2: (a) EDANet [11]. (b) The proposed EDA-FSS. The numbers next to each block denote the feature map size ratios to the input image size. B. Degressive Dilation Block The dilated convolution is widely used by a lot of semantic segmentation networks [3,20]. Typically, these architectures leverage the dilated convolution with incremental dilation rates to enlarge the receptive field in a gradual manner. EDANet a… view at source ↗
Figure 3
Figure 3. Figure 3: Samples of ITRI dataset. (a) RGB image. (b) Ground truth map [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Network A. (b) Network B. The numbers next to each block denote the feature map size ratios to the input image size. C. Feature Size Selection We evaluate the performance of the proposed EDA-FSS on the ITRI dataset. We also design and evaluate two network variants of EDA-FSS called Network A and Network B for comparison. Their architectures are shown in [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sample visual results produced by ERFNet [15], EDANet [11], Network A, Network B, and the proposed EDA-FSS. TABLE III. Description of the structures of each variant. The numbers denote the dilation rates of each EDA module. Only EDA Block 2 and extra block are shown since the rest of these networks are identical. Network EDA￾w/o-di EDA￾DDB-L EDA￾Large-1 EDA￾Large-16 EDA Block 2 1 2 2 2 1 2 2 2 1 4 4 4 1 4 … view at source ↗
Figure 6
Figure 6. Figure 6: Sample visual results produced by the proposed EDA-DDB. (a) RGB input. (b) Ground truth. (c) EDA-DBB. the best performance, which indicates leveraging the decreasing dilation rates in a proper way is still able to lead to improvements. The lane is a thin object, so it would not benefit from a too deep structure. This is one of the reasons that EDA-DDB can surpass EDA-DDB-L though it has fewer parameters. I… view at source ↗
read the original abstract

Lane detection plays an important role in a self-driving vehicle. Several studies leverage a semantic segmentation network to extract robust lane features, but few of them can distinguish different types of lanes. In this paper, we focus on the problem of multi-class lane semantic segmentation. Based on the observation that the lane is a small-size and narrow-width object in a road scene image, we propose two techniques, Feature Size Selection (FSS) and Degressive Dilation Block (DD Block). The FSS allows a network to extract thin lane features using appropriate feature sizes. To acquire fine-grained spatial information, the DD Block is made of a series of dilated convolutions with degressive dilation rates. Experimental results show that the proposed techniques provide obvious improvement in accuracy, while they achieve the same or faster inference speed compared to the baseline system, and can run at real-time on high-resolution images.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The paper proposes two techniques—Feature Size Selection (FSS) and Degressive Dilation Block (DD Block)—for multi-class lane semantic segmentation. Motivated by the observation that lanes appear as small-size, narrow-width objects, FSS selects appropriate feature sizes for thin lane extraction and DD Block applies dilated convolutions with degressive rates to capture fine-grained spatial information. The central claim is that these yield obvious accuracy gains (e.g., mIoU per class) at equal or better inference speed (FPS) versus baselines and support real-time operation on high-resolution images.

Significance. If the reported gains hold under the quantitative tables, the work supplies a practical engineering contribution to efficient multi-class lane segmentation for autonomous driving. The explicit focus on both accuracy and speed, together with the geometry-motivated design choices, addresses a deployment-relevant constraint; the presence of direct baseline comparisons on mIoU and FPS strengthens the empirical case.

minor comments (2)
  1. [Abstract] Abstract: the claim of accuracy gains and speed parity is stated without naming datasets, metrics, or baseline architectures, forcing readers to consult later sections for the quantitative support.
  2. The justification for degressive dilation rates in the DD Block rests on the lane-geometry observation; a short ablation or sensitivity analysis on the rate schedule would strengthen attribution of the reported mIoU gains.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work and the recommendation for minor revision. No specific major comments appear in the provided report, so we have no individual points requiring point-by-point rebuttal at this time. We will address any minor suggestions during revision.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper is an empirical engineering contribution. It motivates FSS and DD Block from the geometric observation that lanes are small and narrow in road images, then reports accuracy and FPS gains via direct table comparisons against baselines on standard datasets. No equations, fitted parameters renamed as predictions, self-citation chains, or ansatzes reduce any claimed result to its own inputs by construction. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper rests on a single domain observation about lane geometry and standard CNN building blocks; no free parameters, invented entities, or additional axioms are stated in the abstract.

axioms (1)
  • domain assumption Lanes appear as small-size and narrow-width objects in typical road-scene images
    This observation directly motivates the introduction of FSS and the DD Block.

pith-pipeline@v0.9.0 · 5686 in / 1180 out tokens · 30586 ms · 2026-05-24T18:05:56.779430+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Real time detection of lane markers in urban streets,

    M. Aly, “Real time detection of lane markers in urban streets,” in IEEE Intelligent Vehicles Symposium, 2008

    work page 2008

  2. [2]

    Segnet: A deep convolutional encoder-decoder architecture for image segmentation,

    V. Badrinarayanan, A. Kendall, and R. Cipolla, “Segnet: A deep convolutional encoder-decoder architecture for image segmentation,” in IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017

    work page 2017

  3. [3]

    Deeplab: Semantic imae segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “Deeplab: Semantic imae segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,” in IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016

    work page 2016

  4. [4]

    Encoder- decoder with atrous separable convolution for semantic image segmentation,

    L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder- decoder with atrous separable convolution for semantic image segmentation,” in European Conference on Computer Vision, 2018

    work page 2018

  5. [5]

    Real-time road scene segmentation with application to lane- mark detection,

    P.-R. Chen, “Real-time road scene segmentation with application to lane- mark detection,” Master thesis, National Chiao Tung University, 2018

    work page 2018

  6. [6]

    Efficient road lane marking detection with deep learning,

    P.-R. Chen, S.-Y. Lo, H.-M. Hang, S.-W. Chan, and J.-J. Lin, “Efficient road lane marking detection with deep learning,” in IEEE International Conference on Digital Signal Processing, 2018

    work page 2018

  7. [7]

    Rich feature hierarchies for accurate object detection and semantic segmentation,

    R. Girshick, J. Donahue, T . Darrell, and J. Malik, “Rich feature hierarchies for accurate object detection and semantic segmentation,” in IEEE Conference on Computer Vision and Pattern Recognition, 2014

    work page 2014

  8. [8]

    Effective u se of dilated convolutions for segmenting small object instances in remote sensing imagery,

    R. Hamaguchi, A. Fujita, K. Nemoto, T. Imaizumi, and S. Hikosaka, “Effective u se of dilated convolutions for segmenting small object instances in remote sensing imagery,” in IEEE Winter Conference on Applications of Computer Vision, 2018

    work page 2018

  9. [9]

    Adam: A method for stochastic optimization,

    D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in International Conference on Learning Representation, 2015

    work page 2015

  10. [10]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G.E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Conference on Neural Information Processing Systems, 2012

    work page 2012

  11. [11]

    Efficient dense modules of asymmetric convolution for real-time semantic segmentation,

    S.-Y. Lo, H. -M. Hang, S. -W. Chan, and J. -J. Lin , “Efficient dense modules of asymmetric convolution for real-time semantic segmentation,” in arXiv preprint arXiv: 1809.06323, 2018

    work page arXiv 2018

  12. [12]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in IEEE Conference on Com puter Vision and Pattern Recognition, 2015

    work page 2015

  13. [13]

    Simple robust road lane detection algorithm ,

    C. Y. Low , H. Zamzuri, and S. A. Mazlan , “ Simple robust road lane detection algorithm ,” i n International Conference on Intelligent and Advanced Systems, 2014

    work page 2014

  14. [14]

    Robust lane detection in shadows and low illumination conditions using local gradient features,

    A. Parajuli, M. Celenk, and H. B. Riley, “Robust lane detection in shadows and low illumination conditions using local gradient features,” in Open Journal of Applied Sciences, 2013

    work page 2013

  15. [15]

    Efficient convnet for real -time semantic segmentation,

    E. Romera, J. M. Alvarez, L. M. Bergasa, and R. Arroyo, “Efficient convnet for real -time semantic segmentation,” in IEEE Intell igent Vehicles Symposium, 2017

    work page 2017

  16. [16]

    Improved river flow and random sample consensus for curve lane detection,

    H. Tan, Y. Zhou, Y. Zhu, D. Yao, and J. Wang, “Improved river flow and random sample consensus for curve lane detection,” in Advances in Mechanical Engineering, 2015

    work page 2015

  17. [17]

    Real-time lane detection by using multiple cues,

    Z. Teng, J.-H. Kim, and D. -J. Kang, “Real-time lane detection by using multiple cues,” in International Conference on Control, Automation and Systems, 2010

    work page 2010

  18. [18]

    Lane detection and tracking using b- snake,

    Y. Wang, E. K. Teoh, and D. Shen, “Lane detection and tracking using b- snake,” in Image and Vision Computing, 2004

    work page 2004

  19. [19]

    Gradient -enhancing conversion for illumination-robust lane detection,

    H. Yoo, U. Yang, and K.Sohn, “Gradient -enhancing conversion for illumination-robust lane detection,” in IEEE Transactions on Intelligent Transportation Systems, 2013

    work page 2013

  20. [20]

    Multi -scale context aggregation by dilated convolutions,

    F. Yu and V. Koltun, “Multi -scale context aggregation by dilated convolutions,” in International Conference on Learning Representation, 2016

    work page 2016

  21. [21]

    Traffic lane detection using fully convolutional neural network,

    J. Zang, W. Zhou, G. Zhang, and Z. Duan, “Traffic lane detection using fully convolutional neural network,” in APSIPA Annual Summit and Conference, 2018

    work page 2018

  22. [22]

    Icn et for real-time semantic segmentation on high -resolution images,

    H. Zhao, X. Qi, X. Shen, J. Shi, and J. Jia, “Icn et for real-time semantic segmentation on high -resolution images,” in European Conference on Computer Vision, 2018

    work page 2018