HSCM-Lane improves lane segmentation with multi-level context modeling

A compact lane-segmentation model enhanced with shifted-window context modeling outperformed a matched ResNet18 baseline on the BDD100K validation set. HSCM-Lane-S, which uses a ResNet18 encoder, reached 34.44% lane-class intersection over union (IoU), 65.00% lane recall and 82.12% balanced accuracy. Without the Hierarchical Swin Context Module (HSCM), the corresponding results were 33.00%, 62.48% and 80.38%. The differences amounted to gains of 1.44 percentage points in IoU, 2.52 points in recall and 1.74 points in balanced accuracy; the IoU increase was equivalent to a 4.36% relative improvement over the no-HSCM counterpart.

Lane segmentation assigns a binary lane-mask label to individual pixels, producing dense geometric information that can support lane keeping, road-structure inference and local trajectory planning in advanced driver-assistance systems and autonomous vehicles. The researchers designed HSCM-Lane for difficult visual conditions, including sparse, thin, low-contrast, occluded and discontinuous lane markings, while retaining a relatively lightweight architecture. Its central feature is the HSCM, which applies shifted-window context modeling at three encoder resolutions: 1/4, 1/8 and 1/16. The resulting context-enhanced features pass through subsequent residual stages and are supplied to a lightweight decoder, where projected static-sum fusion combines them for the final pixel-wise prediction.

The study evaluated three capacities of the same basic architecture. HSCM-Lane-S uses ResNet18, HSCM-Lane-M uses ResNet34 and HSCM-Lane-L uses ResNet50. The HSCM placement, window size, depth, fusion strategy and Focal-Tversky training objective remained fixed, allowing backbone capacity to be compared separately from the context module. Increasing the backbone produced only gradual changes in measured quality: the medium and large configurations achieved lane IoU scores of 34.86% and 34.98%, respectively, compared with 34.44% for the small model. These results show a limited quality-cost trade-off under the reported BDD100K validation protocol rather than a large separation among the three capacities.

The researchers also tested the models under a derived pixel-level out-of-domain evaluation protocol on TuSimple, without using target-domain fine-tuning. HSCM-Lane-S achieved 27.2% lane IoU, while the medium and large variants each reached 27.6%. This experiment indicates that all three configurations transferred to the target dataset under the specified protocol, although the relatively small differences between them do not establish that additional backbone capacity consistently improves out-of-domain performance. The results are based on validation and derived evaluation protocols and do not yet demonstrate benefits for temporal lane tracking, instance-level lane estimation, multimodal sensing, vehicle planning or safety outcomes.

Overall, the reported comparisons associate multi-level shifted-window context modeling with better pixel-wise lane segmentation than the matched architecture without HSCM. The improvement was observed across IoU, recall and balanced accuracy, while the larger ResNet backbones offered only modest changes in validation IoU. The work also maintains a focus on computational economy through a lightweight decoder and offers evidence from a second dataset. Future validation is needed in the areas identified by the authors: temporal and instance-level analysis, multimodal inputs, downstream planning and safety assessment.

Subject of Research: Multi-level shifted-window context modeling for pixel-wise road-lane segmentation

Article Title: HSCM-Lane: A ResNet-based encoder-decoder architecture with multi-level shifted-window context modeling for pixel-wise lane segmentation

Article References: Ton, Q. T., Vu, T. H., Vu, T. N., & Le, T. (2026). HSCM-Lane: A ResNet-based encoder-decoder architecture with multi-level shifted-window context modeling for pixel-wise lane segmentation. PLOS One, 21(10), e0359959. https://doi.org/10.1371/journal.pone.0359959

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0359959

Keywords: lane segmentation, computer vision, shifted-window attention, ResNet, autonomous vehicles, road marking detection, BDD100K, TuSimple, deep learning, HSCM-Lane, ResNet-based, encoder-decoder