Pith. sign in

REVIEW 5 major objections 5 minor 33 references

TraversRL shows that treating sidewalk mapping as a traveler's step-by-step decisions, trained with reinforcement learning on graph-level and stepwise rewards, yields more connected and accurate pedestrian networks than segmentation-based e

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 17:49 UTC pith:L62DWKWQ

load-bearing objection Solid, honest adaptation of iterative graph tracing to pedestrian pathways; the RL fine-tuning story is convincing, but the headline Tile2Net comparison is weakened by a privileged start-node prior that the paper itself concedes. the 5 major comments →

arxiv 2607.17479 v1 pith:L62DWKWQ submitted 2026-07-20 cs.CV

TraversRL: Traversable Pedestrian Pathway Generation With Reinforcement Learning

Bin Han , Robert Wolfe , Bill Howe This is my paper
classification cs.CV
keywords pedestrian pathway extractionreinforcement learningaerial image analysisiterative graph generationconnectivityroutabilitysemantic segmentationgraph generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

TraversRL aims to establish that pedestrian pathway networks—sidewalks, crossings, and footpaths—can be generated directly as connected, routable graphs from aerial images by an agent that grows the network step by step, rather than by segmenting pixels and post-processing the mask. The authors show that a vision-conditioned policy, trained first by supervised imitation and then fine-tuned with reinforcement learning on a terminal graph-level reward (buffered geometry overlap) plus a local stepwise geometric reward, consistently outperforms a state-of-the-art segmentation baseline across three visual backbones and three intersection datasets: on average it improves buffered IoU by 43.6% and a connectivity (traversal-similarity) metric by 205%, with additional gains from the local reward and cleaner, sparser graphs. These results hold zero-shot on unseen cities, and the gains are largest under strict 1-meter alignment. A sympathetic reader would care because accurate, connected sidewalk maps are foundational for accessible navigation and urban planning, and current segmentation pipelines tend to produce fragmented, unroutable networks.

Core claim

The central claim, in the authors' terms, is that iterative graph construction plus outcome-level reinforcement learning is a natural and effective fit for pedestrian pathway extraction. TraversRL encodes the state as a local RGB crop centered on the current node plus a single-channel canvas of previously generated edges, and predicts a discrete direction–distance action (36 angle bins and 10 distance bins, plus a Stop action) that extends the partial graph. The policy is pretrained to imitate ground-truth expansions with localization noise for drift robustness, then fine-tuned with group-relative policy optimization using a terminal buffered-IoU reward (the 'global' variant) or that reward

What carries the argument

The load-bearing mechanism is the vision-conditioned iterative graph-growth policy: at each step the model's state is the concatenation of a local image crop and a canvas rasterizing the partial graph, and the action space is a discretized direction–distance set (361 actions) designed to represent short and long pedestrian segments, with a Stop action to terminate a node's expansion. A supervised pretraining stage teaches the policy to imitate ground-truth edge continuations while injecting localization noise, and a reinforcement learning stage (group-relative policy optimization with KL regularization to the frozen pretrained policy) optimizes the final graph: a terminal buffered-geometry I

Load-bearing premise

The evaluation assumes every rollout starts from a node taken from the ground-truth graph; in deployment, no such node is available, so the claimed reliable routable networks depend on an unprovided seeding mechanism that could shrink the connectivity gains.

What would settle it

Evaluate TraversRL on the same datasets but initialize rollouts from automatic seed proposals (for example, local maxima of a segmentation probability map, or grid points) instead of ground-truth nodes, and recompute traversal similarity and buffered IoU; if the large connectivity advantage over the segmentation baseline mostly disappears, the central contribution is shown to hinge on the ground-truth seeding prior rather than on the learned growth and reward mechanism.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Automatic, routing-ready sidewalk maps could be produced from aerial imagery without the fragmented graphs typical of segmentation-then-vectorization pipelines.
  • Adding a local stepwise geometric reward to a global graph-level reward improves both IoU and traversal similarity while reducing redundant edges, showing that fine-grained credit assignment matters even when the final objective is global.
  • The approach generalizes zero-shot to cities not seen in training, suggesting the learned growth behavior transfers across urban environments.
  • The method's gains are largest under strict geometric alignment (1 m buffer), meaning it improves precise edge placement rather than only coarse overlap.
  • RL fine-tuning consistently yields cleaner, sparser graphs (lower average node degree) while preserving or improving connectivity, so the policy learns to prune noisy branches rather than just maximize overlap.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the method's practical viability is to replace ground-truth start-node seeding with an automatic seed proposer (e.g., from a segmentation prior) and measure how the reported connectivity gains change; the paper's own conclusion flags this as an open requirement.
  • The stepwise overlap-efficiency reward is a generic credit-assignment signal that could be transferred to other sequential graph-generation tasks with non-differentiable terminal objectives, such as road or river network tracing.
  • Because a single rollout can only recover the connected component reachable from its seed, multi-seed inference with a seed proposal/filtering strategy is a natural next step toward 'walk the city' de novo mapping, and its effectiveness is testable on scenes with multiple disconnected components.
  • The fixed normalized crop size creates a scale ambiguity across datasets with different ground resolutions; making the observation scale-aware is a concrete, testable modification that could further close the zero-shot transfer gap.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces TraversRL, a vision-conditioned iterative graph generator for pedestrian pathway extraction from aerial imagery. The method grows a graph step-by-step from a seed node using a discretized direction–distance action space, trained first by supervised imitation (TraversRL-Pre) and then by RL fine-tuning with either a global graph-level reward (TraversRL-Global) or an additional per-step local geometric reward (TraversRL-Local). Experiments across three backbones (ResNet-18, ViT-S, Swin-T) and three datasets (WashingtonInter, PathwayBench Seattle, PathwayBench D.C.) report large improvements over the Tile2Net segmentation baseline in buffered IoU and TravSim, with RL fine-tuning giving further consistent gains and cleaner graphs. The paper also presents buffer-size and action-space ablations, and reports zero-shot transfer to unseen cities.

Significance. If the reported results are reliable, the paper makes a useful contribution to an understudied problem: pedestrian pathway extraction that explicitly targets connectivity and routability, not just pixel-level segmentation. The internal comparison across three backbones and three datasets, with seed-to-seed standard deviations reported, is a strength, as is the focus on a traversal-oriented evaluation metric (TravSim) in addition to IoU. The idea of modeling pathway generation as sequential decision making with RL fine-tuning is well motivated and the local stepwise reward is a reasonable, clearly specified mechanism. However, the headline external comparison to Tile2Net rests on an asymmetric evaluation protocol that gives TraversRL access to ground-truth start nodes while giving the segmentation baseline no equivalent prior. This is acknowledged in Section 6 but not resolved, and it is load-bearing for the paper's central claim.

major comments (5)
  1. [Sec. 4.3 / Sec. 6] The evaluation protocol gives TraversRL a privileged start-node prior. Section 4.3 states 'We evaluate each method over five randomly sampled start nodes,' and Section 3.2 says that for training and evaluation the start node is 'a random node in the ground-truth graph'; Section 6 concedes that in deployment this information is unavailable and a seeding mechanism is required. Tile2Net is given no such anchor: it must propose a complete network from the image, whereas TraversRL only needs to grow the connected component reachable from a known pathway point. This biases both headline metrics: IoU benefits because new edges radiate from a location guaranteed to lie on the pathway, and TravSim benefits because the seed is already connected to the ground-truth topology. The reported +205% TravSim gain over Tile2Net could shrink substantially under a fair comparison. The paper should either eva
  2. [Sec. 4.2] The Tile2Net baseline's training condition is not specified. It is unclear whether Tile2Net was retrained on the WashingtonInter training splits, fine-tuned on PathwayBench, or used as an off-the-shelf frozen model. Since the paper emphasizes zero-shot transfer and reports absolute IoU/TravSim numbers for Tile2Net, the reader cannot assess whether the comparison is apples-to-apples. If Tile2Net was not retrained on the same data, its lower performance may reflect domain shift or resolution mismatch rather than a genuine limitation of segmentation-based methods. This directly affects the paper's central claim that TraversRL 'substantially improves buffered IoU ... relative to a state-of-the-art segmentation baseline.' Please provide the exact checkpoints, training/validation protocols, and preprocessing used for Tile2Net.
  3. [Sec. 3.3 / Sec. 4.3] The buffered IoU metric used for evaluation is also the terminal reward for the RL fine-tuning (Sec. 3.3: 'terminal, graph-level reward R_i is buffered IoU in our experiments'; Sec. 4.3 uses a 1-meter buffer for evaluation). Consequently, the reported IoU gains of TraversRL-Global over TraversRL-Pre are partly the result of directly optimizing the evaluation metric, and are not evidence that the model improves an independent measure of geometric quality. The paper should state this circularity explicitly and report a metric not used as reward — TravSim is one such metric, and the paper does report it, which mitigates the concern, but the IoU gains should be framed as optimization outcomes rather than intrinsic improvements. The buffer-size ablation (Sec. 5.4) is interesting but reinforces this point by showing that gains shrink at wider buffers.
  4. [Sec. 5.1 / Table 1] Statistical significance of the headline gains is not established beyond small seed standard deviations. For example, Table 1 reports Tile2Net with std 0.046 IoU on WashingtonInter, while TraversRL-Pre values are far higher, so the difference is clearly large there; however, on Seattle IoU the gap between Tile2Net (0.169±0.021) and TraversRL-Pre (e.g., 0.186±0.003 for ResNet-18) is much smaller relative to the reported variance, and no test accounts for the fact that only five random starts are used. Since the central external claim depends on these numbers, a test or at least a bootstrap confidence interval over start-node choices should be reported.
  5. [Sec. 6] The paper itself lists 'Coverage of disconnected components' as a limitation: a single-seed rollout can only recover the component reachable from that seed. This is not merely a deployment concern; it is also an evaluation concern for the in-domain WashingtonInter scores, since the dataset construction 'discard[s] scenes whose ground-truth graph is disconnected' (Sec. 4.1). This filtering removes the hardest cases for the method, and makes the reported numbers less representative of real-world scenes. The paper should report results on unfiltered data or at least quantify the fraction of real scenes excluded by this criterion.
minor comments (5)
  1. [Fig. 1 caption] Typo: 'preductions' should be 'predictions.'
  2. [Sec. 4.1] The PathwayBench datasets are cited as [32], but one of the authors of the current paper is also a co-author of PathwayBench. This is not disclosed in the main text; a brief acknowledgment of the potential overlap would be helpful.
  3. [Sec. 4.2] For VecRoad and NETracer, the paper says these road-oriented baselines 'do not transfer well,' but the reported numbers are extremely low (e.g., <0.001 IoU for VecRoad), which suggests the conversion pipeline may not be well tuned. The authors should provide more detail on how baselines were configured so readers can judge whether the comparison is fair.
  4. [Sec. 5.5] The action-space ablation uses only ResNet-18 and TraversRL-Local. This is acceptable but should be stated more prominently as a single-backbone result, since the paper elsewhere emphasizes multi-backbone consistency.
  5. [General] No code or trained models are provided for reproducibility. Given the number of rollout hyperparameters (top-k, restarts, stop threshold, max expansions, KL coefficient, local reward weight), a public release would be important for independent verification.

Circularity Check

0 steps flagged

No derivation-level circularity; the headline comparison is empirical and externally benchmarked. Minor caveats: the RL reward equals the buffered-IoU evaluation metric, and TravSim comes from a benchmark co-authored by a current author.

full rationale

The paper's central claim is an empirical comparison of TraversRL against Tile2Net on held-out intersection datasets; it is not derived from a fitted parameter or from a self-citation chain. The RL reward in Eq. (3) is buffered IoU, the same quantity as the Sec. 4.3 evaluation metric, which makes the RL-fine-tuning IoU gains partly an optimization outcome rather than an independent discovery; however, this does not make the Tile2Net comparison circular because the baseline is not trained with that reward and the test sets are held out. TravSim is taken from PathwayBench [32], co-authored by one of the current authors, but it is an external benchmark and is not used to justify a uniqueness claim or to forbid alternatives. The ground-truth start-node initialization is a real methodological limitation, explicitly acknowledged in Sec. 6, but it is a fairness/validity concern rather than a circularity. No equation in the paper reduces to its own input by construction, and no prediction is statistically forced by a fitted parameter renamed as a result.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method rests on standard ML design parameters plus domain assumptions about annotation quality, seeding, and metric validity. No new physical or ontological entities are introduced. The most consequential free design choices are the action-space grid, the RL hyperparameters, and the rollout configuration, all chosen by hand or by validation.

free parameters (6)
  • Action-space bin counts (Nθ=36, Nr=10) = 36 angle bins, 10 distance bins, +Stop = 361 actions
    Chosen by hand as the default; Sec. 5.5 (T5) shows results are sensitive to this discretization, and larger action spaces improve in-domain IoU but hurt zero-shot transfer on Seattle.
  • Localization noise scale η = 0.02 (normalized coordinates)
    Added in supervised pretraining (Sec. 3.1, Eq. 2) to mitigate rollout drift; magnitude chosen by hand.
  • KL coefficient β = 0.05
    KL regularization weight in the GRPO objective (Eq. 4); set during RL fine-tuning.
  • Local reward shaping weight λ = 0.1, with stepwise rewards clipped to [-3,3]
    Controls the balance between terminal advantage and stepwise reward in Eq. (7); selected by validation.
  • Buffer radius r for reward = 0.01 normalized (≈1 m)
    Used for buffered IoU reward; the same 1 m radius defines the headline evaluation metric.
  • Rollout hyperparameters (top-k, restarts R, stop threshold τ_stop, max expansions M) = Tuned per backbone on validation set
    Sec. 3.2: these are selected to maximize validation buffered IoU, making the evaluation metric part of configuration selection.
axioms (5)
  • domain assumption Ground-truth pathway annotations are complete and accurate enough to serve as learning targets and evaluation labels.
    Dataset construction (Sec. 4.1) involves OSM-derived boxes, an internal annotation database, and manual inspection, but errors in the labels propagate into both the reward and the reported metrics.
  • ad hoc to paper A start node from the ground-truth graph is available at rollout time.
    Sec. 4.3 samples random start nodes from ground truth; Sec. 6.2 acknowledges deployment lacks this information, so the evaluation is conditional on a privileged seed.
  • domain assumption The discretized direction-distance action space is expressive enough to represent the true graph geometry within tolerance.
    Only 10 distance bins over [0,0.2] normalized coordinates are used; the T5 ablation shows performance varies with angle granularity, so the default grid is load-bearing.
  • domain assumption Buffered IoU with a 1 m radius and TravSim on clipped tiles are valid proxies for navigation quality.
    These metrics define both the reward and the evaluation (Sec. 4.3), but the paper does not validate them against human mobility outcomes, such as actual wheelchair routes or wayfinding times.
  • domain assumption The local state (image crop plus generation canvas) is a sufficient representation for sequential expansion decisions.
    The policy in Sec. 3.1 conditions on a 0.5×0.5 window and a rasterized canvas; larger-scale context, such as global intersection layout, is not directly modeled.

pith-pipeline@v1.3.0-alltime-deepseek · 12925 in / 10703 out tokens · 102078 ms · 2026-08-01T17:49:37.202674+00:00 · methodology

0 comments
read the original abstract

Automatically generating pedestrian pathways from aerial images requires producing a connected network suitable for routing, not just detecting where sidewalks appear. Sidewalks and crossings, in contrast to roads, may be partially occluded, implicitly defined, and exhibit complex connectivity patterns. Existing segmentation-based approaches focus on labeling pixels to infer segments, but often produce disconnected or fragmentary graphs that are unreliable for navigation. We introduce TraversRL, a vision-conditioned model that iteratively grows a pathway network from an aerial image, simulating a traveler navigating the built environment. TraversRL uses an action space of short and long direction-distance segments designed to adapt to complex patterns and span occlusions, and uses a combination of graph-level and step-wise rewards to balance overall connectivity with precise edge placement. Across three visual backbones and three intersection datasets, TraversRL substantially improves buffered IoU with the ground-truth graph relative to a state-of-the-art segmentation baseline, and more than doubles metrics of connectivity. Moreover, combining global and local rewards produces cleaner graphs with fewer spurious branches while further improving overall performance. These results demonstrate that modeling pathway extraction as a sequential decision process from the perspective of a traveler, while optimizing for final graph quality with reinforcement learning, produces significantly more reliable pedestrian networks.

Figures

Figures reproduced from arXiv: 2607.17479 by Bill Howe, Bin Han, Robert Wolfe.

Figure 1
Figure 1. Figure 1: Qualitative examples of pedestrian pathway generation. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Supervised node expansion in TraversRL pretraining. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Local stepwise reward. We compare the buffered ground-truth region G (blue) with the current predicted buffer Pt−1 (gray). Adding a new edge at step t contributes newly added buffered area ∆At (orange), of which the overlapped portion with ground-truth is ∆It. We define the stepwise reward rt = ∆It/(∆At +ϵ): it is high when the added edge aligns better with ground-truth (good step), moderate under partial … view at source ↗
Figure 4
Figure 4. Figure 4: RL fine-tuning with local stepwise reward shaping (TraversRL [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: RL fine-tuning produces cleaner, sparser graphs. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 12 linked inside Pith

  1. [1]

    Ahmetovic,D.,Manduchi,R.,Coughlan,J.M.,Mascetti,S.:Zebracrossingspotter: Automaticpopulationofspatialdatabasesforincreasedsafetyofblindtravelers.In: Proceedingsofthe17thInternationalACMSIGACCESSConferenceonComputers & Accessibility. pp. 251–258 (2015) 4

    2015

  2. [2]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Bastani, F., He, S., Abbar, S., Alizadeh, M., Balakrishnan, H., Chawla, S., Madden, S., DeWitt, D.: Roadtracer: Automatic extraction of road networks from aerial images. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4720–4728 (2018) 3, 5

    2018

  3. [3]

    arXiv preprint arXiv:2506.07885 (2025) 4

    Bhuyan, Z., Xie, Y., Rith, A., Yan, X., Apostolov, N., Oke, J., Ai, C.: Crosswalknet: An optimized deep learning framework for pedestrian crosswalk detection in aerial images with high-performance computing. arXiv preprint arXiv:2506.07885 (2025) 4

    Pith/arXiv arXiv 2025

  4. [4]

    ISPRS Journal of Photogrammetry and Remote Sensing228, 741–755 (2025) 3

    Deng, L., Deng, Y., Meng, Y., Chen, J., Xi, Z., Liu, D., Chu, Q.: Gld-road: A global–local decoding road network extraction model for remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing228, 741–755 (2025) 3

    2025

  5. [5]

    arXiv preprint arXiv:2010.11929 (2020) 10

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 10

    Pith/arXiv arXiv 2010

  6. [6]

    arXiv preprint arXiv:2410.15115 (2024) 2, 4

    Gao, J., Xu, S., Ye, W., Liu, W., He, C., Fu, W., Mei, Z., Wang, G., Wu, Y.: On designing effective rl reward at training time for llm reasoning. arXiv preprint arXiv:2410.15115 (2024) 2, 4

    Pith/arXiv arXiv 2024

  7. [7]

    arXiv preprint arXiv:2501.12948 (2025) 2, 4

    Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., et al.: Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025) 2, 4

    Pith/arXiv arXiv 2025

  8. [8]

    In: Proceedings of the 27th annual ACM symposium on User interface software and technology

    Hara, K., Sun, J., Moore, R., Jacobs, D., Froehlich, J.: Tohme: detecting curb ramps in google street view using crowdsourcing, computer vision, and machine learning. In: Proceedings of the 27th annual ACM symposium on User interface software and technology. pp. 189–204 (2014) 4

    2014

  9. [9]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016) 10

    2016

  10. [10]

    In: European Conference on Computer Vision

    He,S.,Bastani,F.,Jagwani,S.,Alizadeh,M.,Balakrishnan,H.,Chawla,S.,Elshrif, M.M., Madden, S., Sadeghi, M.A.: Sat2graph: Road graph extraction through graph-tensor encoding. In: European Conference on Computer Vision. pp. 51–67. Springer (2020) 3

    2020

  11. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Hetang, C., Xue, H., Le, C., Yue, T., Wang, W., He, Y.: Segment anything model for road network graph extraction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2556–2566 (2024) 3

    2024

  12. [12]

    Computers, Environment and Urban Systems101, 101950 (2023) 4, 10

    Hosseini, M., Sevtsuk, A., Miranda, F., Cesar Jr, R.M., Silva, C.T.: Mapping the walk: A scalable computer vision approach for generating sidewalk network datasets from aerial imagery. Computers, Environment and Urban Systems101, 101950 (2023) 4, 10

    2023

  13. [13]

    arXiv preprint arXiv:2412.16720 (2024) 2, 4

    Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Carney, A., et al.: Openai o1 system card. arXiv preprint arXiv:2412.16720 (2024) 2, 4

    Pith/arXiv arXiv 2024

  14. [14]

    International Journal of Geographical Information Science 27(5), 947–962 (2013) 4 ECCV 2026 Submission #none 17

    Karimi,H.A.,Kasemsuppakorn,P.:Pedestriannetworkmapgenerationapproaches and recommendation. International Journal of Geographical Information Science 27(5), 947–962 (2013) 4 ECCV 2026 Submission #none 17

    2013

  15. [15]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Liu,C.,Jiang,Y.,Zheng,N.:Netracer:Atopology-awareiterativetracingapproach for tubular structure extraction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 20593–20602 (2025) 3, 10

    2025

  16. [16]

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchical visiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021) 10

    2021

  17. [17]

    ISPRS Journal of Photogrammetry and Remote Sensing175, 340–352 (2021) 3

    Lu, X., Zhong, Y., Zheng, Z., Zhang, L.: Gamsnet: Globally aware road detection network with multi-scale residual learning. ISPRS Journal of Photogrammetry and Remote Sensing175, 340–352 (2021) 3

    2021

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Mi, L., Zhao, H., Nash, C., Jin, X., Gao, J., Sun, C., Schmid, C., Shavit, N., Chai, Y., Anguelov, D.: Hdmapgen: A hierarchical graph generative model of high definition maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4227–4236 (2021) 3

    2021

  19. [19]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    O’Meara, J.S., Hwang, J., Wang, Z., Saugstad, M., Froehlich, J.E.: Rampnet: A two-stage pipeline for bootstrapping curb ramp detection in streetscape images from open government metadata. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6656–6665 (2025) 4

    2025

  20. [20]

    arXiv preprint arXiv:1707.06347 (2017) 4

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) 4

    Pith/arXiv arXiv 2017

  21. [21]

    URL https://arxiv

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv. org/abs/2402.033002(3), 5 (2024) 2, 4

    Pith/arXiv arXiv 2024

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Sun, T., Di, Z., Che, P., Liu, C., Wang, Y.: Leveraging crowdsourced gps data for road extraction from aerial imagery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7509–7518 (2019) 3

    2019

  23. [23]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Tan, Y.Q., Gao, S.H., Li, X.Y., Cheng, M.M., Ren, B.: Vecroad: Point-based iterative graph exploration for road graphs extraction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8910– 8918 (2020) 3, 10

    2020

  24. [24]

    5: Scaling reinforcement learning with llms

    Team, K., Du, A., Gao, B., Xing, B., Jiang, C., Chen, C., Li, C., Xiao, C., Du, C., Liao, C., et al.: Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599 (2025) 2, 4

    Pith/arXiv arXiv 2025

  25. [25]

    Transportation research record2678(7), 845–856 (2024) 4

    Verma, R., Ukkusuri, S.V.: Crosswalk detection from satellite imagery for pedes- trian network completion. Transportation research record2678(7), 845–856 (2024) 4

    2024

  26. [26]

    arXiv preprint arXiv:2504.20571 (2025) 2, 4

    Wang, Y., Yang, Q., Zeng, Z., Ren, L., Liu, L., Peng, B., Cheng, H., He, X., Wang, K., Gao, J., et al.: Reinforcement learning for reasoning in large language models with one training example. arXiv preprint arXiv:2504.20571 (2025) 2, 4

    Pith/arXiv arXiv 2025

  27. [27]

    arXiv preprint arXiv:2506.14245 (2025) 2, 4

    Wen,X.,Liu,Z.,Zheng,S.,Ye,S.,Wu,Z.,Wang,Y.,Xu,Z.,Liang,X.,Li,J.,Miao, Z., et al.: Reinforcement learning with verifiable rewards implicitly incentivizes correct reasoning in base llms. arXiv preprint arXiv:2506.14245 (2025) 2, 4

    Pith/arXiv arXiv 2025

  28. [28]

    IEEE Transactions on Geoscience and Remote Sensing61, 1– 14 (2023) 3

    Xu, Q., Long, C., Yu, L., Zhang, C.: Road extraction with satellite images and partial road maps. IEEE Transactions on Geoscience and Remote Sensing61, 1– 14 (2023) 3

    2023

  29. [29]

    IEEE Transactions on Geoscience and Remote Sensing60, 1–12 (2022) 3 18 B

    Xu, Z., Liu, Y., Gan, L., Sun, Y., Wu, X., Liu, M., Wang, L.: Rngdet: Road network graph detection by transformer in aerial images. IEEE Transactions on Geoscience and Remote Sensing60, 1–12 (2022) 3 18 B. Han, R. Wolfe, B. Howe

    2022

  30. [30]

    IEEE Robotics and Automation Letters8(5), 2991–2998 (2023) 3

    Xu, Z., Liu, Y., Sun, Y., Liu, M., Wang, L.: Rngdet++: Road network graph detection by transformer with instance segmentation and multi-scale features en- hancement. IEEE Robotics and Automation Letters8(5), 2991–2998 (2023) 3

    2023

  31. [31]

    arXiv preprint arXiv:2303.02323 (2023) 1, 4

    Zhang, Y., Bolten, N., Mehta, S., Caspi, A.: Ape: An open and shared an- notated dataset for learning urban pedestrian path networks. arXiv preprint arXiv:2303.02323 (2023) 1, 4

    Pith/arXiv arXiv 2023

  32. [32]

    arXiv preprint arXiv:2407.16875 (2024) 4, 10, 11

    Zhang, Y., Howe, B., Mehta, S., Bolten, N.J., Caspi, A.: Pathwaybench: Assessing routability of pedestrian pathway networks inferred from multi-city imagery. arXiv preprint arXiv:2407.16875 (2024) 4, 10, 11

    Pith/arXiv arXiv 2024

  33. [33]

    ISPRS International Journal of Geo-Information10(1), 39 (2021) 3

    Zhou, K., Xie, Y., Gao, Z., Miao, F., Zhang, L.: Funet: A novel road extraction net- work with fusion of location data and remote sensing imagery. ISPRS International Journal of Geo-Information10(1), 39 (2021) 3

    2021