SC-Lane: Slope-Aware 3D Lane Detection

• SC-Lane is a slope-aware and temporally consistent road height estimation framework that improves camera-based 3D lane detection on sloped, curved, and uneven roads.
• It employs an adaptive multi-slope fusion module (SAAF) with ground-consistency supervision to accurately model road surfaces as a geometric prior.
• Temporal consistency during training refines height predictions, yielding state-of-the-art results with a 64.3% F-score and lower height estimation errors on OpenLane.

SC-Lane denotes the “Slope-aware and Consistent Road Height Estimation Framework for 3D Lane Detection,” a camera-based 3D lane detection method that treats road height estimation as a first-class geometric prior rather than an auxiliary by-product of lane prediction. Its central claim is that accurate reconstruction of the road’s height profile improves 3D lane localization, particularly on sloped roads, curves, merges/splits, intersections, and uneven terrain. The framework introduces a Slope-Aware Adaptive Feature module for adaptive multi-slope fusion and a Height Consistency Module for temporal regularization, and on OpenLane it reports a 64.3% F-score together with substantial gains on dedicated road-height metrics (Park et al., 14 Aug 2025).

1. Definition and terminological scope

In the provided literature, the name SC-Lane is explicitly defined as a slope-aware and temporally consistent heightmap estimation framework for 3D lane detection (Park et al., 14 Aug 2025). The acronym expands to the two properties that organize the method: slope-aware adaptive fusion of height features and consistent temporal regularization of the estimated road surface.

The term is not used uniformly across adjacent lane-perception literature. In the ETSC-based ego-lane identification paper, the text explicitly states that “SC-Lane” is not explicitly introduced as an acronym there, and that the safest interpretation is that it refers informally to the paper’s lane-identification pipeline rather than to a separately defined method (Jin et al., 2024). This distinction matters because SC-Lane, in its formal sense, is about road height estimation for 3D lane detection, not about ego-lane hypothesis selection in an HD map.

A common misconception is therefore to treat SC-Lane as a generic label for any structured or sequential lane method. The provided corpus does not support that interpretation. Within these sources, SC-Lane is specifically the height-estimation framework of the 2025 paper, whereas related works address different subproblems such as 2D lane detection, lane continuity modeling, lane topology reasoning, multi-frame segmentation, or ego-lane identification.

2. Problem setting and motivation

SC-Lane addresses camera-based 3D lane detection under the premise that lane geometry cannot be recovered reliably unless the road surface is modeled accurately (Park et al., 14 Aug 2025). The paper argues that if the estimated road surface is wrong, lane geometry becomes distorted, lane points may be projected to incorrect 3D positions, BEV representations become less reliable, and performance degrades especially on sloped roads, curves, merges/splits, intersections, and uneven terrain.

The method is positioned against two broad families of prior 3D lane methods. BEV-based methods project image features to BEV using IPM or variants of it; the cited examples are BEVLaneDet, LaneCPP, PVALane, and HeightLane. Their limitations, as stated in the source, are that BEV projection is sensitive to camera extrinsic errors and that such methods often struggle with complex slopes and changing road geometry. Transformer-based methods, including LATR, PETRv2, and RFTR, operate directly in front-view image space using 3D positional encodings, but they do not explicitly solve road height estimation.

The immediate precursor singled out by SC-Lane is HeightLane, which introduced multi-slope anchors but fuses them uniformly. The source identifies two specific shortcomings: it treats all slope features equally, and it lacks explicit temporal consistency across frames. SC-Lane is designed to address exactly these two issues. This suggests that the paper’s contribution is not merely a stronger lane head, but a reformulation of height estimation as a structured, scene-dependent, and temporally coherent inference problem.

3. Architectural formulation

The full SC-Lane pipeline consists of eight stages: input image, backbone feature extraction, multi-slope anchor projection, Slope-Aware Adaptive Feature, height decoder, Height Consistency Module, height-guided spatial transform, and 3D lane head (Park et al., 14 Aug 2025). The backbone is ResNet-50.

The method begins from a multi-slope anchor formulation. Each anchor $a \in \{1,\dots,A\}$ is defined at BEV location $(X,Y)$ with fixed height $Z_a$, i.e. as a 3D point $(X,Y,Z_a)$. It is projected to the image plane and sampled from the image feature map: $$F^{a}_{sampled} = \mathcal{S}(\mathcal{P}(X, Y, Z_a), F_{img}).$$ This produces slope-specific features for each BEV location.

The Slope-Aware Adaptive Feature (SAAF) module replaces uniform anchor fusion with learned, image-conditioned weighting. A lightweight CNN upsamples backbone features to BEV resolution, predicts an $A$-channel tensor, and applies softmax across anchors so that

$\sum_{a=1}^{A}\alpha_a = 1.$

The fused height-aware feature is then

$$F_{SA}(X,Y)=\sum_{a=1}^{A}\alpha_a(X,Y)\,F^{a}_{sampled}(X,Y).$$

According to the paper’s interpretation, flatter anchors dominate on flat roads, steeper anchors receive higher weight on uphill or downhill scenes, and multiple anchors can be blended in ambiguous regions. The important point is that slope modeling becomes adaptive and spatially varying, rather than fixed across the scene.

SC-Lane also imposes ground-consistency supervision on the adaptive weights. The anchor weights are reinterpreted as a confidence heightmap: $$\mathcal{H}_{conf}(X,Y)=\sum_{a=1}^{A}\alpha_a(X,Y)\,\text{Anchor}_a(X,Y), \qquad \text{Anchor}_a(X,Y)=Z_a.$$ This confidence heightmap is projected back into the image plane and compared with a binary ground mask through the auxiliary loss

$$\mathcal{L}_{SA} = \text{IoU}(\mathbf{M}_{ground}, \mathbf{I}\cdot\mathbf{E}\cdot T_{\text{road}\to\text{ego}}\cdot \mathcal{H}_{conf}).$$

The stated role of this term is to suppress implausible slope selection and concentrate anchor weights on valid road regions.

The second major component is the Height Consistency Module (HCM), which regularizes heightmaps across consecutive frames. The previous frame’s heightmap is aligned to the current frame by ego-motion compensation: $(X,Y)$0 with

$(X,Y)$1

Consistency is then enforced on the overlapping region $(X,Y)$2: $(X,Y)$3 The paper motivates this specifically by the relative staticness of the road surface after ego-motion compensation.

The total height-estimation objective is

$(X,Y)$4

with $(X,Y)$5, $(X,Y)$6, and $(X,Y)$7. The resulting heightmap is then used in a height-guided spatial transform that refines BEV features before final prediction by the 3D lane head.

4. Heightmap representation, training protocol, and evaluation

SC-Lane defines the BEV heightmap at 200 × 48 resolution with 0.5 m per pixel, corresponding to the prediction range used for 3D lane detection (Park et al., 14 Aug 2025). For OpenLane, the heightmaps are generated by accumulating ground-surface 3D point clouds using LiDAR data and segmentation labels. For Apollo Synthetic, where LiDAR is unavailable, heightmaps are derived from depth information, and only frames with semantic annotations are used.

A notable methodological contribution is the paper’s insistence on standardized road-height metrics. It evaluates height estimation with Mean Absolute Error (MAE),

$(X,Y)$8

Root Mean Squared Error (RMSE),

$(X,Y)$9

and threshold-based accuracy

$Z_a$0

with $Z_a$1. The source explicitly notes that these metrics are common in surface and depth estimation but have been underutilized for road height assessment.

The implementation uses a ResNet-50 backbone and 600 × 800 input images. Training is performed on 4 A6000 GPUs. The joint height-estimation-plus-lane-detection regime runs for 24 epochs with batch size 8, while height estimation alone is trained for 16 epochs with batch size 16. These details are important because SC-Lane’s temporal component is a supervision mechanism rather than a recurrent inference-time module.

5. Quantitative results

On OpenLane validation, SC-Lane reports an F-score of 64.3%, compared with 50.5 for PersFormer, 53.1 for Anchor3DLane, 58.4 for BEV-LaneDet, 60.3 for LaneCPP, 61.9 for LATR, and 62.7 for both HeightLane and PVALane (Park et al., 14 Aug 2025). The source characterizes this as state-of-the-art 3D lane detection performance.

Scenario-wise results are also reported. SC-Lane attains 70.7 on curved roads, 63.0 on merge/split, 56.1 at intersections, 54.6 on up/down terrain, 58.1 in extreme weather, and 56.5 at night. The paper explicitly interprets its strongest relative behavior as occurring on curved roads, merge/split scenes, and intersections, which is consistent with the claim that adaptive slope fusion is especially useful when road geometry changes rapidly.

For road-height estimation, the method improves both absolute error and thresholded accuracy:

Metric	HeightLane	SC-Lane
MAE	0.235	0.176
RMSE	0.343	0.259
[email protected]	0.253	0.293
[email protected]	0.442	0.507
[email protected]	0.673	0.756

The same section reports a Single Anchor baseline at 0.234 MAE, 0.367 RMSE, 0.253 [email protected], 0.452 [email protected], and 0.688 [email protected], indicating that the gains are not solely due to the presence of a height branch but to the adaptive multi-slope formulation itself.

The ablation study further decomposes the contribution of each module. Starting from Baseline (HeightLane) at 62.7 F-score, adding Height Consistency raises performance to 63.2, adding SAA Feature* to 63.9, adding SAA Feature to 64.1, and combining all components yields 64.3. The corresponding height metrics improve in parallel. The source concludes that SAAF provides the largest single gain, that ground-mask supervision further improves SAAF, and that Height Consistency alone helps modestly but complements SAAF.

SC-Lane is also compared with multi-frame 3D lane methods. Anchor3DLane with 2 frames reports 54.3, Curveformer++ reports 55.4 with 2 frames and 53.5 with 3 frames, and GTANet with 2 frames reports 62.4, whereas SC-Lane reports 64.3 with 2 frames’ temporal cue classified as Supervision rather than Input (Park et al., 14 Aug 2025). The paper emphasizes that temporal consistency is used during training only, so inference remains single-frame and incurs no extra runtime input cost.

6. Relationship to neighboring research areas

SC-Lane belongs to the broader lane-perception literature but occupies a distinct niche. It is neither a generic 2D lane detector nor a topology-reasoning system. CLRNet focuses on cross-layer refinement, ROIGather, and Line IoU (LIoU) for 2D lane detection (N et al., 2023). RCLane models each lane as a chain of relay stations and uses segmentation, transfer maps, and distance maps to decode lanes with complex topologies such as Y-shape, fork-shape, and nearly horizontal lanes (Xu et al., 2022). ContinuityLearner emphasizes geometric continuity in lane segmentation by combining lane-pixel coordinates with lane-specific visual features in a point-set completion framework (Fang et al., 2021). UniTopo unifies lane detection and lane-topology prediction by representing lane-to-lane topology as connected lanes within a shared perception pipeline (Li et al., 9 May 2026). A separate sequential lane detector uses multi-frame images with spatial-temporal attention and LSTM to improve robustness in occlusion, dazzle lighting, shadows, blur, and mixed traffic scenes (Patil et al., 3 Feb 2026).

These distinctions clarify what SC-Lane does and does not do. It does not primarily solve lane instance segmentation, lane continuity completion, lane topology reasoning, or sequential mask prediction. Its distinctive object is the road heightmap, which it uses as a geometric prior to improve downstream 3D lane detection. A plausible implication is that SC-Lane is best understood as a geometric prior learning framework within the 3D lane stack, complementary to detectors that focus on proposal structure, topology, or temporal appearance aggregation.

The distinction from ego-lane identification is especially sharp. The ETSC-based method in “Towards Accurate Ego-lane Identification with Early Time Series Classification” models online evidence accumulation over measurement model quality (MMQ) time series to decide which mapped lane currently contains the ego vehicle, reporting 99.6% accuracy with 0.84 s average prediction time (Jin et al., 2024). That task concerns lane membership within an HD map, whereas SC-Lane concerns height-aware 3D lane reconstruction from camera input.

7. Limitations and implications

The limitations stated or implied in the source are specific. First, the Height Consistency Module is used only in training, so SC-Lane does not directly exploit temporal information during inference (Park et al., 14 Aug 2025). Second, the heightmap representation is tailored to structured road scenes; the text notes that generalization to unstructured terrain may require adaptation. Third, the source observes that Z-error is not always the best reflection of performance for dense BEV heightmap methods, which indicates that evaluation for road geometry remains only partially standardized.

At the same time, the paper argues that its metric suite—MAE, RMSE, and [email protected]/0.1/0.2—establishes a more rigorous standard for road-height evaluation than was previously common. It also suggests broader uses for road-height estimation, including HD map reconstruction, localization, and autonomous-driving perception more broadly. This suggests a broader research implication: if road height is a stable and structured signal, then explicit geometric supervision may be as consequential for 3D lane understanding as architectural innovations in lane-head design.

Within the provided literature, SC-Lane therefore represents a specific shift in emphasis. Rather than asking only how to detect lanes, it asks how to recover the road surface on which lanes lie, and how adaptive slope modeling plus temporal consistency can make that surface reliable enough to improve 3D lane detection under complex driving conditions (Park et al., 14 Aug 2025).