Source-linked AI summary
RoadTracer: Automatic Extraction of Road Networks from Aerial Images
Favyen Bastani, Songtao He, Sofiane Abbar, Mohammad Alizadeh, Hari Balakrishnan, Sanjay Chawla, Sam Madden, David DeWitt
TL;DR
Mapping road networks is expensive and labor-intensive, while segmentation-based methods struggle to infer connectivity from noisy local predictions. RoadTracer directly constructs road graphs through CNN-guided iterative search, capturing 45% more junctions than segmentation at a 5% average error rate across 15 cities.
Problem
Road mapping remains expensive and labor-intensive, and segmentation methods delegate road connectivity to error-prone heuristic post-processing of noisy local CNN outputs.
Method
RoadTracer iteratively constructs a road-network graph by adding road segments one at a time, using a CNN decision function that receives aerial imagery and the graph built so far.
Results
At a 5% average error rate on a junction-by-junction matching metric, RoadTracer captured 45% more junctions than the segmentation approach across 15 cities, 0.58 versus 0.40.
Takeaways & Limitations
The results suggest RoadTracer is an important step toward fully automating map construction from aerial images, where accurate junction topology is crucial for navigation.
Takeaways & Limitations
CNN training on near-perfect graph inputs causes inference errors to compound when the constructed graph deviates from ground truth.
Abstract
from arXiv · showhide
Mapping road networks is currently both expensive and labor-intensive. High-resolution aerial imagery provides a promising avenue to automatically infer a road network. Prior work uses convolutional neural networks (CNNs) to detect which pixels belong to a road (segmentation), and then uses complex post-processing heuristics to infer graph connectivity. We show that these segmentation methods have high error rates because noisy CNN outputs are difficult to correct. We propose RoadTracer, a new method to automatically construct accurate road network maps from aerial images. RoadTracer uses an iterative search process guided by a CNN-based decision function to derive the road network graph directly from the output of the CNN. We compare our approach with a segmentation method on fifteen cities, and find that at a 5% error rate, RoadTracer correctly captures 45% more junctions across these cities.
1. Introduction
Road mapping remains costly and error-prone because occlusions and noisy pixel-level segmentation make road connectivity difficult to recover. RoadTracer instead constructs the road graph iteratively with CNN guidance and captures more junctions than a segmentation approach.
- Motivation: Creating and updating road maps is tedious, expensive, and often manual, despite their importance for location-based services and autonomous vehicles.Map providers still receive many tens of thousands of error reports per day.
- Motivation: Occlusions from trees, buildings, and shadows make road-network connectivity difficult to infer from aerial imagery, even for humans.These conditions complicate automatic extraction from images.
- Prior limitation: Segmentation methods classify pixels locally, then rely on noisy outputs and complex heuristic post-processing to infer graph connectivity.The CNN is not trained to decide how road segments inter-connect.
- RoadTracer: RoadTracer iteratively adds individual road segments, using a CNN decision function conditioned on the partial network constructed so far.This removes the intermediate road-image representation and reduces reliance on extensive post-processing.
- Evaluation: 45% more junctions were correctly captured by RoadTracer than by the segmentation approach at a 5% average error rate across 15 cities.The corresponding junction-capture values were 0.58 for RoadTracer and 0.40 for the segmentation approach.
2. Related Work
Prior road-mapping methods either segment pixels and heuristically reconstruct graphs or directly use detailed models and assumptions. These approaches remain vulnerable to uncertainty, complex topology, and incomplete coverage.
- Segmentation approaches: Road segmentation estimates whether each pixel belongs to a road, but its probabilistic output cannot directly serve as a road-network graph.CNNs and other probabilistic models are commonly used for this pixel-classification task.
- Segmentation approaches: Thresholding, morphological thinning, and centerline tracing convert segmentation outputs into graphs, while DeepRoadMapper adds heuristic stages for missing connections.These methods perform best when road segmentation has modest error.
- Limitations: Heuristics perform poorly when segmentation is uncertain because of occlusion, ambiguous topology, or parallel and multi-layer roads.These cases challenge graph reconstruction from pixel-level predictions.
- Direct graph extraction: Other direct-image methods use complex road models, contextual assumptions, shape-based footprints, or road-tree growth.Their procedures avoid standard segmentation but still involve numerous heuristics and assumptions.
- Alternative data sources: GPS trajectories can improve mapping where imagery hides roads or layered interchanges, but collecting sufficient spatial and temporal coverage is challenging.GPS data may also support more frequent map updates.
3. Automatic Map Inference
Segmentation-based map inference converts pixel predictions into graphs through heuristic post-processing, but extensive noise can leave disconnected or inaccurate maps. RoadTracer instead constructs the graph iteratively with a CNN-guided search that directly adds road segments and uses the partial graph as context.
- Prior Work: Segmentation Approaches: Segmentation-based methods label pixels as road or non-road, then convert the output into a graph through thresholding, thinning, and simplification.Morphological thinning produces one-pixel-wide centerlines, which are interpreted as graph vertices and edges before Douglas-Peucker simplification.
- Prior Work: Segmentation Approaches: Refinement heuristics prune short dangling segments, extend dead ends, and merge nearby junctions to remove basic segmentation noise.These operations yield a refined graph, but they address only basic noise types.
- Prior Work: Segmentation Approaches: Many gaps in noisy segmentation outputs can produce disconnected graphs with poor road coverage despite refinement.The paper presents cases where the noise is too extensive for heuristic correction.
- RoadTracer: Iterative Graph Construction: RoadTracer searches from a known road location, maintaining a graph and vertex stack while a CNN decision function chooses the next action.At each step, the search either adds a road segment or stops and returns to the preceding vertex.
- RoadTracer: Iterative Graph Construction: The search adds vertices and edges by walking a fixed distance D = 12 meters in a selected direction, or backtracking when the decision is to stop.Directions are selected from uniformly distributed angles, and the search stops when no unexplored adjacent roads remain.
- RoadTracer: Iterative Graph Construction: RoadTracer supplies the partial graph and local aerial imagery to its CNN, enabling direct graph construction with only thresholding and maximum-index selection as post-processing.The graph channel records explored roads and provides context for interpreting occluded regions.
4. Iterative Graph Construction CNN Training
RoadTracer trains a CNN decision function within an iterative graph search rather than relying on static oracle-generated states. Dynamic training exposes the CNN to its own imperfect partial graphs, while map matching helps select relevant ground-truth paths for labeling.
- 4.1. Static Training Dataset: Static oracle-generated examples produce nearly perfect partial graphs, unlike inference states where CNN decisions introduce errors that compound into larger prediction errors.The CNN is trained on states constructed from ground truth, so it does not learn to recover from its own imperfect inputs.
- 4.1. Static Training Dataset: Static training can cause a small directional error at a turn to make the search veer off the actual road and fail to correct course.Adding artificial noise reduces the problem but does not match the noise produced inherently by the CNN during inference.
- 4.2. Dynamic Labels: RoadTracer dynamically generates training examples by running the search with the evolving CNN as its decision function.As the model changes, the generated examples change as well, keeping training states aligned with the model’s current behavior.
- 4.2. Dynamic Labels: During dynamic training, the CNN updates the partial graph and vertex stack, while an oracle supplies labels for back-propagation rather than controlling the search.This separates state generation from label computation and exposes the CNN to states induced by its own outputs.
- 4.2. Dynamic Labels: Map matching selects a ground-truth path most similar to a randomly walked path in the current graph before deriving unexplored-road labels.The procedure uses a configurable walk length of w = 10 and avoids reusing already explored ground-truth edges.
5. Evaluation
The evaluation compares RoadTracer with segmentation-based baselines on aerial imagery from cities held out from training, using junction, topology, and shortest-path metrics. RoadTracer generally offers a better accuracy–error tradeoff, especially at practical error rates, while standard topology and path metrics may reward noisy maps.
- Dataset: The dataset uses aerial imagery and OSM road graphs for 40 cities, with 25 cities for training and 15 separate cities for testing.Each city covers approximately 24 square kilometers, and the imagery has 60 cm/pixel resolution.
- Metrics and baselines: The evaluation compares RoadTracer with DeepRoadMapper and a segmentation-based approach using TOPO, SP, and a new junction metric.The junction metric compares inferred and ground-truth junctions by captured and erroneous incident edges.
- Metric limitations: TOPO and SP can assign higher scores to noisier maps and make the causes of scores difficult to interpret, motivating the junction metric.The proposed metric aims to represent practical usability while remaining interpretable.
- Quantitative results: The junction metric reports 45% better average recall for RoadTracer than segmentation at 5% error, with Fcorrect values of 0.58 versus 0.40.RoadTracer has a better Ferror for a given Fcorrect, with the largest improvement between 5% and 10% error.
- Quantitative results: On TOPO, RoadTracer has lower error than segmentation when recall is below 0.43, while segmentation surpasses it only above 0.14 error, where maps are reported unusable.Above 0.43 recall, lowering RoadTracer's walk-stop threshold provides marginal recall gains but substantially increases error.
- Quantitative results: RoadTracer outperforms segmentation on shortest-path evaluation because segmentation noise can make inferred shortest paths much shorter than ground-truth paths.SP results are reported at thresholds yielding the highest number of correct shortest paths.
- Qualitative results: DeepRoadMapper performs poorly on this dataset, while qualitative results show RoadTracer handling frequent building and shadow occlusions better in Chicago and Boston.Performance is comparable in the Salt Lake City and Toronto examples, and DeepRoadMapper produces many disconnections.
6. Conclusion
RoadTracer directly constructs road-network graphs with a CNN-guided iterative search, avoiding segmentation intermediates and post-processing heuristics. Across 15 cities, it captures 45% more junctions than the segmentation approach at a 5% error rate.
- RoadTracer uses a CNN-guided iterative graph construction method to output road networks directly from aerial images.The method adds road segments one at a time based on the partial network constructed so far.
- 45% more junctions were correctly captured by RoadTracer than by the segmentation approach at a 5% error rate, with scores of 0.58 versus 0.40.