DeH4R: A Decoupled and Hybrid Method for Road Network Graph Extraction

Abstract

The automated extraction of complete and precise road network graphs from remote sensing imagery remains a critical challenge in geospatial computer vision. Segmentation-based approaches, while effective in pixel-level recognition, struggle to maintain topology fidelity after vectorization postprocessing. Graph-growing methods build more topologically faithful graphs but suffer from computationally prohibitive iterative ROI cropping. Graph-generating methods first predict global static candidate road network vertices, and then infer possible edges between vertices. They achieve fast topology-aware inference, but limits the dynamic insertion of vertices. To address these challenges, we propose DeH4R, a novel hybrid model that combines graph-generating efficiency and graph-growing dynamics. This is achieved by decoupling the task into candidate vertex detection, adjacent vertex prediction, initial graph contruction, and graph expansion. This architectural innovation enables dynamic vertex (edge) insertions while retaining fast inference speed and enhancing both topology fidelity and spatial consistency. Comprehensive evaluations on CityScale and SpaceNet benchmarks demonstrate state-of-the-art (SOTA) performance. DeH4R outperforms the prior SOTA graph-growing method RNGDet++ by 4.62 APLS and 10.18 IoU on CityScale, while being approximately 10 $\times$ faster. The code will be made publicly available at https://github.com/7777777FAN/DeH4R.

Figures (7)

• Figure 1: Pipelines of existing methods and our method. (a) Graph-growing methods are computationally expensive due to its iterative ROI cropping and full-model forward passes. (b) Graph-generating methods show great potential in parallel inference, but existing methods infer edges between a fixed set of candidate vertices (or candidate vertices and their adjacent vertices extracted in one-step), precluding dynamics for vertex expansion. (c) Our approach decouples the task into candidate vertex detection, adjacent vertex prediction, initial graph construction and graph expansion, thus enabling dynamics and high inference efficiency.
• Figure 2: The pipeline of DeH4R. The upper row depicts the complete workflow of DeH4R, aligned with our decoupling strategy: Given an input image, CVD extracts features and detects candidate vertices, AVP predicts adjacent vertices for all candidates, IGC converts vertex predictions into a graph (inference) or generates labels for predicted vertices on-the-fly (training), and GEP expands the graph $N$ times to obtain a complete one. The lower row provides detailed explanations of CVD, AVP, IGC and GEP.
• Figure 3: Schematic diagram of the effect before and after decoding. (a) A discrepancy between candidate vertices and predicted adjacent vertices before decoding; (b) The connected candidate vertices after decoding. Yellow points are candidate vertices. Red point is the currently processing vertex. Dashed and white points and lines are predicted adjacent vertices and possible edges, respectively. Orange lines are confirmed edges.
• Figure 4: Examples of graph expansion. (a) An insertion case where the predicted vertex is not in existing graph; (b) A merging case where the predicted vertex is close to an existing vertex and then merged.
• Figure 5: A schematic connecting current candidate vertex $v_n$ to another candidate $v_c$ (both from the detected candidate vertex set $C$), based on their respective adjacency predictions $u_n$ and $u_c$. The radius threshold $r$ defines the search range for locating an existing candidate vertex (e.g., $v_c$) that may correspond to a predicted adjacent vertex (e.g., $u_c$ corresponding to $v_n$, or $u_n$ corresponding to $v_c$). The predicted edges are denoted as $e_1 = {\{v_n, u_n\}}$ and $e_2 = {\{v_c, u_c\}}$, while $e = {\{v_n, v_c\}}$ represents the candidate edge to be verified. $s_1$ and $s_2$ denote the Euclidean distances between the possible vertex pairs ($v_n, u_c$) and ($v_c, u_n$), respectively. $\theta$ is the angle discrepancy between the predicted edges ($e_1$/$e_2$) and the candidate edge $e$.