Refine Connections, Close the Gap: A Reliable Enhancement Framework for Driving Scene Topology

Xiaoqi Wang; Dingyi Zhaung; David Paz; Wenbin He; Yucai Bai; Peng Zhou; Rui Zhang; Jinhua Zhao

cs.CVOct 8, 2026

Refine Connections, Close the Gap: A Reliable Enhancement Framework for Driving Scene Topology

Authors: Xiaoqi Wang, Dingyi Zhaung, David Paz, Wenbin He, Yucai Bai, Peng Zhou, Rui Zhang, Jinhua Zhao, +1 more

Organizations: Bosch Research North America · Bosch Center for Artificial Intelligence (BCAI) · Massachusetts Institute of Technology · Bosch XC China

Abstract

In autonomous driving, understanding scene topology - the connectivity between lanes and traffic elements - is critical for safe path planning and motion control. While current methods excel at detecting individual map elements, their connectivity reasoning often falls short of its theoretical potential, leaving a significant performance gap relative to the theoretical upper-bound achievable given the underlying detections. Furthermore, the decision-ready topology graphs passed to downstream tasks often remain unreliable. Current approaches typically derive connectivity by thresholding continuous topology scores; however, these scores often fail to reflect the true logical likelihood of connectivity, resulting in false positives or missing connections. Existing benchmarks further overlook this issue by primarily evaluating continuous metrics, rather than assessing the discrete connectivity required for decision-making. To bridge these gaps, we propose TopoEnhance, a novel topology enhancement framework designed to unlock the latent potential of existing methods and improve the reliability of decision-ready topology. We formulate topology enhancement as a denoising-based reconstruction process, where the model learns to recover structural consistency from stochastically corrupted ground-truth graphs. This formulation enables the model to resolve logical inconsistencies and rectify unreliable connections, producing robust discrete topology graphs that closely approach theoretical maximum performance. Extensive experiments across different baselines show that TopoEnhance consistently improves both continuous topology metrics (TOP score), and discrete connectivity measured by our adapted Topology Jaccard Similarity (TJS) metric. As a flexible, source-agnostic framework, TopoEnhance delivers substantial gains across diverse state-of-the-art baselines without requiring retraining.

Figures & tables

Figure 1: Overall framework of TopoEnhance. Training phase: (a) Forward process (topology noise simulation): a novel noise simulation strategy that stochastically corrupts ground-truth topology graphs into noisy graphs $\mathcal{G}_{\mathrm{aug}}$ by injecting realistic geometric and structural prediction errors. (b) Reverse process (topology denoising): a lightweight heterogeneous GNN learns to recover reliable topology connections from the simulated noisy graphs. Inference phase: once trained, the enhancement model is directly applied to calibrate topology scores from any topology reasoning method, without retraining.

Dataset	Method	Before TopoEnhance		After TopoEnhance		Upper-Bound
Dataset	Method	TOP ${}_{ll}\uparrow$	TOP ${}_{lt}\uparrow$	TOP ${}_{ll}\uparrow$	TOP ${}_{lt}\uparrow$	TOP ${}_{ll}\uparrow$	TOP ${}_{lt}\uparrow$
Argoverse2 (Subset_A)	TopoNet ( Li et al., 2026 )	10.9	23.8	21.8	25.8	24.4	28.8
	TopoMLP ( Wu et al., 2024 )	21.6	26.9	24.4	28.7	25.0	31.4
	Topo2D ( Li et al., 2024 )	22.3	26.2	24.3	27.3	25.0	29.8
	TopoLogic ( Fu et al., 2024 )	23.9	25.4	24.5	27.2	26.0	30.6
	SMART ( Ye et al., 2025 )	37.0	33.0	40.1	35.4	40.9	38.9

Table 1: Enhanced results on OpenLane-V2 measured by the TOP metric (version 2.1.0), which evaluates the quality of continuous topology scores. Bolded blue values indicate improved score over the corresponding baselines before applying TopoEnhance.

Dataset	Method	Before TopoEnhance		After TopoEnhance		Upper-bound
Dataset	Method	TJS ll $\uparrow$	TJS lt $\uparrow$	TJS ll $\uparrow$	TJS lt $\uparrow$	TJS ll $\uparrow$	TJS lt $\uparrow$
Argoverse2 (Subset_A)	TopoNet ( Li et al., 2026 )	16.3	31.0	32.9	55.8	44.3	63.7
	TopoMLP ( Wu et al., 2024 )	18.5	31.5	59.7	59.8	69.7	72.8
	Topo2D ( Li et al., 2024 )	18.9	34.5	55.0	51.5	68.7	70.8
	TopoLogic ( Fu et al., 2024 )	21.0	32.5	43.9	53.3	50.9	71.4
	SMART ( Ye et al., 2025 )	38.1	36.6	87.7	82.8	92.9	90.1

Table 2: Enhanced results on OpenLane-V2 evaluated using the TJS metric, which measures the quality of discrete topology connections.

Panel of Figure 2 — Figure 2: Qualitative comparison of topology connections before and after enhancement.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Method	TOP ll	TOP lt	TJS ll	TJS lt
TopoNet	3.34	14.40	11.84	20.49
+ TopoEnhance	10.23	16.37	12.87	35.49
Upper-Bound	11.47	18.42	16.90	37.37

Appendix

Table 3: Evaluation on the geographically non-overlapping Argoverse2 GeoSplits. Both TopoNet and TopoEnhance are trained on Argoverse2 GeoSplits.

Dataset	Method	DET l	DET t	TOP ll	TOP lt	TJS ll	TJS lt
Argoverse2 (Subset_A)	TopoPoint, reported	31.40	55.30	28.70	30.00	–	–
	TopoPoint, our reproduction	28.82	54.20	28.10	30.36	22.31	35.94
	+ TopoEnhance	28.82	54.20	28.21	32.28	25.03	55.13
	Upper-Bound	28.82	54.20	29.65	36.58	31.66	56.67
nuScenes (Subset_B)	TopoPoint, reported	31.20	60.20	28.30	27.10	–	–
	TopoPoint, our reproduction	23.67	60.66	24.73	22.42	14.09	35.13

Appendix

Table 4: Additional evaluation of TopoEnhance on TopoPoint. “Reported” denotes the results reported in the TopoPoint paper, while “our reproduction” uses checkpoints trained by us with the official implementation. Since TJS is not reported by TopoPoint, the corresponding entries are omitted.

Dataset	Method	Detection		Before TopoEnhance			After TopoEnhance			Upper-Bound
Dataset	Method	DET ${}_{l}\uparrow$	DET ${}_{t}\uparrow$	TOP ${}_{ll}\uparrow$	TOP ${}_{lt}\uparrow$	OLS $\uparrow$	TOP ${}_{ll}\uparrow$	TOP ${}_{lt}\uparrow$	OLS $\uparrow$	TOP ${}_{ll}\uparrow$	TOP ${}_{lt}\uparrow$	OLS $\uparrow$
Argoverse2 (Subset_A)	TopoNet ( Li et al., 2026 )	28.6	48.6	10.9	23.8	39.8	21.8	25.8	43.7	24.4	28.8	45.1
	TopoMLP ( Wu et al., 2024 )	28.5	49.5	21.6	26.9	44.1	24.4	28.7	45.2	25.0	31.4	46.0
	Topo2D ( Li et al., 2024 )	29.1	50.6	22.3	26.2	44.5	24.3	27.3	45.3	25.0	29.8	46.1
	TopoLogic ( Fu et al., 2024 )	29.9	47.2	23.9	25.4	44.1	24.5	27.2	44.7	26.0	30.6	45.9
	SMART ( Ye et al., 2025 )	46.6	47.7	37.0	33.0	53.1	40.1	35.4	54.3	40.9	38.9	55.2

Appendix

Table 5: Complete OpenLane-V2 evaluation corresponding to Table 1 . DET l and DET t measure lane and traffic-element detection, respectively. Since TopoEnhance does not modify detections, these metrics remain unchanged. OLS is computed from the displayed detection and topology metrics following the standard OpenLane-V2 evaluation protocol.

Backbone Model	Dim	$\text{TOP}_{ll}$	$\text{TOP}_{lt}$
DINOv2-ViT-S	384	20.3	24.64
DINOv2-ViT-B	768	21.8	25.06
DINOv2-ViT-L	1024	21.84	25.8
DINOv2-ViT-G	1536	21.84	25.75
DINOv3-ViT-S	384	21.56	24.48
DINOv3-ViT-B	768	21.83	24.35

Appendix

Table 6: Ablation studies on different visual backbones for traffic-element node feature extraction. All experiments are conducted using TopoNet as the baseline on Argoverse2 (Subset_A).

Loss TOP ll TOP lt BCE 12.1 22.0 BCE + Focal 22.2 22.0 Normalized BCE (Ours) 21.8 25.8

Appendix

Table 7: Ablation studies on loss functions. This experiment is conducted to enhance TopoNet’s predictions on Argoverse2 (Subset_A) under a consistent training configuration.

Strategy TOP ll TOP lt TopoEnhance 21.8 25.8 TopoEnhance without Jitter 21.7 25.4

Appendix

Table 8: Ablation studies on geometric localization jitter, while keeping other components from topology noise simulation unchanged. This experiment is conducted to enhance TopoNet’s predictions on Argoverse2 (Subset_A) under a consistent training configuration.

Model	Noise	TOP ll		TOP lt
		Before TopoEnhance	After TopoEnhance	Before TopoEnhance	After TopoEnhance
TopoNet	$p=0.0$	10.90	21.80	23.80	25.80
	$p=0.1$	4.62	9.99	15.81	17.11
	$p=0.2$	2.78	6.14	11.31	12.35
	$p=0.3$	1.99	4.39	8.48	9.39
	$p=0.5$	1.30	2.77	4.94	5.64

Appendix

Table 9: Robustness analysis under varying levels of topology noise on Argoverse2 (Subset_A).

Model	Before TopoEnhance	After TopoEnhance
TopoNet	24.48%	24.21%
TopoMLP	12.62%	7.77%
TopoLogic	10.68%	8.27%
SMART	10.52%	6.38%

Appendix

Table 10: Temporal consistency measured by edge prediction flip rate on Argoverse2 (Subset_A). Lane correspondences across frames are established using persistent lane segment IDs, and only ground-truth lane pairs that remain co-visible for at least

K=5

consecutive frames are evaluated. The flip rate is the fraction of adjacent frame pairs in which the predicted discrete connectivity changes between positive and negative. Lower values indicates that the edge flip less frequently.

Model	Before TopoEnhance	After TopoEnhance
TopoNet	0.166	0.118
TopoMLP	0.111	0.080
TopoLogic	0.285	0.228
SMART	0.104	0.075

Appendix

Table 11: Temporal consistency measured by the standard deviation of topology scores on Argoverse2 (Subset_A). Lane correspondences across frames are established using persistent lane segment IDs, and only ground-truth lane pairs that remain co-visible for at least

K=5

consecutive frames are evaluated. For each lane pair, the temporal standard deviation is computed over its topology scores across consecutive frames. Lower values indicate more stable topology confidence over time.

Model	Relation	Error Type	Total	Corrected	Broken
TopoNet	Lane–Lane	Missed ego successor	243	60	0
	Lane–Lane	Missed intersection turn	5,378	1,502	0
	Lane–Traffic	Missed red light	153	103	0
		Missed green light	398	311	0
		Missed yellow light	54	23	0
		Missed turn signal	49	36	0

Appendix

Table 12: Safety-relevant topology errors corrected and introduced by TopoEnhance on Argoverse2 (Subset_A), grouped into lane–lane and lane–traffic relations. Lane–lane errors include missed ego successors and intersection turns, while lane–traffic errors include missed associations with traffic signals. Only relations whose endpoint detections are successfully matched to ground truth are evaluated. “Corrected” denotes a previously missing ground-truth connection recovered after refinement, while “Broken” denotes an originally correct connection that becomes incorrect.

Model	Before TopoEnhance	After TopoEnhance
TopoNet	39.19%	68.92%
TopoMLP	94.81%	100%
TopoLogic	92.41%	93.67%
SMART	91.36%	98.77%

Appendix

Table 13: Scene-level ego-route recoverability on Argoverse2 (Subset_A). Frame-level topology predictions are aggregated into a scene-level lane graph using persistent Argoverse2 lane IDs. We evaluate only routes for which the required lane detections are successfully detected and matched to ground truth. For each such scene, the ego vehicle’s actual driving route is considered recoverable only if every lane transition along the complete route is correctly recovered in the predicted scene-level topology graph. Higher ego-route recoverability indicates that a larger fraction of scenes have their complete ego trajectories successfully recovered.

Component	Latency (ms)	Percentage
Feature Extraction	0.21	3.2%
GNN Message Passing	5.21	80.9%
Score Prediction	0.64	9.9%
Total	6.43	100%

Appendix

Table 14: Latency breakdown of TopoEnhance on a representative OpenLane-V2 scene (200 lanes, 100 TEs).

Component	Memory Overhead
Feature Extraction	+0.1 MB
GNN Message Passing	+0.0 MB
Score Prediction	+0.2 MB
Total	+0.3 MB

Appendix

Table 15: Memory overhead of each inference component.

Panel of Figure 3 — Figure 3: Qualitative results for scene 10023: predictions before and after refinement with TopoEnhance across baselines.

Panel of Figure 4 — Figure 4: Qualitative results for scene 10013: predictions before and after refinement with TopoEnhance across baselines (SMART, Topo2D, TopoLogic, TopoMLP, TopoNet). All panels correspond to the same ground-truth view (left).

Panel of Figure 5 — Figure 5: Qualitative results for scene 10021: predictions before and after refinement with TopoEnhance across baselines (SMART, Topo2D, TopoLogic, TopoMLP, TopoNet). All panels correspond to the same ground-truth view (left).

Panel of Figure 6 — Figure 6: Qualitative results for scene 10030: predictions before and after refinement with TopoEnhance across baselines.

Explore similar work

Unified Modeling of Lane and Lane Topology for Driving Scene Reasoning
May 9, 2026Han Li, Yulu Gao, Si Liu +3Autonomous Driving Perception
School of Artificial Intelligence, Beihang University, Beijing, China · Zhongguancun Academy, Beijing, China · Hangzhou International Innovation Institute, Beihang University, Hangzhou, China +1
TopoHR: Hierarchical Centerline Representation for Cyclic Topology Reasoning in Driving Scenes with Point-to-Instance Relations
Apr 27, 2026Yifeng Bai, Zhirong Chen, Bo Song +2Hierarchical Representation Learning Autonomous Driving Perception
Institute of Intelligent Machines, HFIPS, Chinese Academy of Sciences · University of Science and Technology of China · NullMax +1
HGeo-TopoMap: Boosting Topological Mapping with Hierarchical Geometric Priors
Jul 23, 2026Siyu Li, Kunyu Peng, Di Wen +3HD Map Construction Autonomous Driving Perception
The authors are with the School of Automation and Electrical Engineering, Zhejiang University of Science and Technology, China. · The authors are with the Institute for Anthropomatics and Robotics, Karlsruhe Institute of Technology, Germany. · The authors are with the School of Artificial Intelligence and Robotics and the National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, China

May 9, 2026, cs.CVJ/K move · Enter open · S save

Unified Modeling of Lane and Lane Topology for Driving Scene Reasoning

Han Li, Yulu Gao, Si Liu, Yuhang Wang+2

School of Artificial Intelligence, Beihang University, Beijing, China · Zhongguancun Academy, Beijing, China · Hangzhou International Innovation Institute, Beihang University, Hangzhou, China+1

Autonomous Driving Perception

Open paper AskPDF arXiv

Autonomous vehicles need to perceive not only physical elements in the driving scene, such as lane lines and traffic lights, but also logical elements like lane centerlines and their topology. Existing lane topology reasoning methods typically follow a reasoning-by-detection paradigm, where lane topological relationships are primarily derived from lane detection results. In this paper, we propose an innovative method called Unified Modeling of Lane and Lane Topology (UniTopo), which represents the topological relationships between lanes as connected lanes, encompassing predecessor lanes, successor lanes, and their interconnections. This unified representation of lanes and lane topology allows us to simultaneously obtain both the positions and topological information of lanes within a shared perception pipeline, establishing a new paradigm for directly perceiving lane topology from original image features. We validate our method on the driving scene reasoning benchmark OpenLane-V2, which consists of two subsets, built based on Argoverse2 and nuScenes, respectively. Our method achieves TOP_ll of 30.1% and 31.8% on the two subsets, significantly surpassing the existing state-of-the-art method T^2SG by 6.0% and 8.6%.

Abstract

Explore similar work

Unified Modeling of Lane and Lane Topology for Driving Scene Reasoning

TopoHR: Hierarchical Centerline Representation for Cyclic Topology Reasoning in Driving Scenes with Point-to-Instance Relations

HGeo-TopoMap: Boosting Topological Mapping with Hierarchical Geometric Priors