acceptodds
Under review as a conference paper at ICLR 2027

VERA-ROAD: VISUAL EVIDENCE REASONING AND ADAPTIVE REPAIR FOR ROAD NETWORK EXTRACTION

Abstract

Road-network extraction from very-high-resolution remote sensing imagery is difficult not only because roads vary across scenes, but also because different regions within the same image fail for different reasons. A narrow road may require stronger geometric continuity, a low-contrast segment may benefit from fine- detail enhancement, and a visually ambiguous junction may require a wider spatial context. Most extraction pipelines nevertheless process all regions through a largely fixed computational path. We introduce VERA-Road, a region-wise visual decision framework that replaces this static processing pattern with an observe–act–verify– repair loop. A shared vision-language policy first interprets local, neighboring, and global evidence and issues an action for each region. The proposed Contextual Action Composer (CAC) realizes the action by combining complementary detail, geometry, and cross-band operators. Rather than assuming that the issued action is always sufficient, Outcome-Guided Repair (OGR) measures the resulting prediction state and selectively applies directional-alignment or multi-scale-context repair only to unresolved regions. Finally, Topology-Rewarded Policy Learning (TRPL) optimizes the action policy from downstream graph quality with a group- relative objective, avoiding the need for region-level action annotations. On the SpaceNet road benchmark, the resulting model reaches 86.79 F1, 93.47 precision, 82.51 recall, and 87.42 APLS, while the ablation and sensitivity studies show complementary gains from action composition, selective repair, and topology- aware policy optimization.

Primary area:
foundation or frontier models, including LLMs
Keywords:
road extraction, remote sensing, deep learning, vision foundation models, large language models, reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_ojw3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper recasts road-network extraction as region-wise visual decision making. On top of a frozen vision encoder and road readout, candidate regions are derived from prediction uncertainty and graph structure. For each region, a vision-language policy (Qwen2.5-VL with LoRA) inspects local, neighboring, and global context together with numerical measurements, and issues intensity levels for three complementary feature operators (Haar-based detail, directional geometry, and multispectral cross-band), which the Contextual Action Composer (CAC) converts into gated residual feature corrections. An Outcome-Guided Repair (OGR) module then measures whether the issued action has resolved the region and selectively triggers one of two repairs (directional re-encoding or multi-scale context re-encoding) for unresolved regions only. The policy is trained with Topology-Rewarded Policy Learning (TRPL), a GRPO-style objective whose reward reflects gains in graph-level metrics (TOPO-F1 and APLS) minus an action-cost penalty, requiring no region-level action annotations.

Critical strengths
  1. The closed-loop observe-act-verify-repair formulation is a sensible way to inject adaptivity into an otherwise static extraction pipeline, and the verify step is what distinguishes it from unconditional feature reprocessing.

  2. To my knowledge this is the first work in road-network extraction that lets a VLM agent decide how each region should be processed and trains it purely from task-level graph rewards, opening a direction beyond static pipelines.

Critical weaknesses
  1. All results come from a single dataset (SpaceNet), and the compared methods are incomplete, with recent road-graph extractors such as SAM-Road, RNGDet++, and TopDiG missing from Table 1.

  2. The experimental section is thin (one main table and one ablation), and there is no analysis of the learned policy's behavior on which the central claim rests.

Questions for authors
  1. Please add SAM-Road (and ideally RNGDet++/TopDiG) under the same split.

  2. Please replace the VLM policy with a small trainable controller under the same GRPO reward. This would directly test whether vision-language reasoning is necessary.

  3. Please provide an analysis of the learned policy, such as action distributions across region types and qualitative examples of the issued actions. This is essential to support the claim of "visual evidence reasoning" in the title, by showing that the selected actions are semantically consistent with the visual evidence.

Reviewer_SrggOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VERA-Road uses a vision-language policy to select regional feature enhancements and selectively repair road predictions. CAC combines detail, geometry, and multispectral operators; OGR selects directional or multi-scale repair; TRPL optimizes the policy using graph-quality rewards. On SpaceNet, the paper reports F1 86.79 and APLS 87.42, supported by staged ablations and a repair-threshold analysis.

Critical strengths
  1. CAC and OGR address heterogeneous local failures through complementary operators and selective repair. Graph-quality rewards align policy learning with road connectivity without requiring manual action labels.

  2. Promising internal gains. Table 2 reports APLS increasing from 82.56 to 87.42 and F1 from 80.73 to 86.79. Subject to evaluation consistency, these gains support the practical potential of the combined framework.

Critical weaknesses
  1. Critical implementation details are missing. Sections 3.1–3.3 do not specify the encoder/readout variants and checkpoints, region-generation algorithm, measurement formulas, operator architectures, prompts, or action-level counts. Figure 3 sweeps verification thresholds but does not identify the main configuration. These omissions prevent reconstruction of core components from the supplied manuscript.

  2. Several reported metrics require reconciliation. Table 1 reports MaGRoad and DeH4R F1 scores of 84.37 and 85.11, above the harmonic means of their displayed precision/recall: 83.82 and 84.29. Ordinary same-weight macro-averaging cannot explain these discrepancies. The metric definitions and aggregation rules must be clarified before the comparison can be trusted.

  3. VERA-Road uses RGB and multispectral observations, but baseline access to these inputs is unspecified and no RGB-only ablation is reported. Its SpaceNet split/resolution also differs from the public DeH4R protocol, without documented matched reruns. These factors confound attribution of the comparative gains.

  4. The strong base obscures improvement attribution. Table 2's unidentified base reaches 82.56 APLS, already 8.00 points above Table 1's strongest comparator. The complete method adds 4.86 points to that base. Without explaining this initial advantage, the reported 12.86-point margin does not isolate the contribution of CAC, OGR, and TRPL.

  5. The VLM controller lacks matched ablations. Table 2 couples added operators with learned routing. Fixed-weight fusion and a lightweight router using comparable observations are needed to isolate the controller's benefit. This matters given total computation of 132.08 versus 17.82 TFLOPs/image for the base (Section 4.5), approximately a 7.41× increase.

  6. Directly relevant baselines are absent. Table 1 omits SAM-Road, RNGDet++, and SAM-Road++, although all are cited. The SAM-based methods are particularly relevant to the frozen feature/readout design, while RNGDet++ is an improved counterpart to the included RNGDet. Matched comparisons would better establish the method's position.

  7. The evidence is limited to one benchmark. All experiments use one SpaceNet split, leaving the stability of learned routing and repair across datasets untested. Although this is an acknowledged scope limitation, additional graph-benchmark evidence would strengthen the claimed general utility.

  8. Prior agentic visual-processing work is insufficiently discussed. Neither RestoreAgent [1] nor AgenticIR [2] is cited, despite their relevance to tool selection and post-execution reflection. The paper should distinguish its regional formulation and topology feedback from these established workflows to clarify the incremental contribution.

[1] Chen H, Li W, Gu J, et al. Restoreagent: Autonomous image restoration agent via multimodal large language models[J]. Advances in Neural Information Processing Systems, 2024, 37: 110643-110666.

[2] Zhu K, Gu J, You Z, et al. An intelligent agentic system for complex image restoration problems[C]//International Conference on Learning Representations. 2025, 2025: 57985-58013.

Questions for authors

Please see the weaknesses.

Reviewer_9PS6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VERA-Road, a region-adaptive road-network extraction framework consisting of Contextual Action Composer (CAC), Outcome-Guided Repair (OGR), and Topology-Rewarded Policy Learning (TRPL). The method reports strong results on SpaceNet, including 86.79 F1 and 87.42 APLS. While the framework is interesting and clearly presented, the current experiments do not sufficiently establish that the gains originate from the proposed reasoning and policy-learning components.

Critical strengths
  1. The observe-act-verify-repair formulation is intuitive and well motivated for heterogeneous road regions.

  2. CAC, OGR, and TRPL form a conceptually coherent framework.

  3. The reported SpaceNet results are strong, particularly the APLS score of 87.42.

Critical weaknesses

1. The base system is insufficiently specified

The paper does not clearly identify the architectures, checkpoints, or training procedures of the frozen visual encoder and road readout. This is especially important because the Base model already achieves 82.56 APLS, substantially exceeding the best baseline APLS of 74.56 in Table 1. Therefore, much of the apparent improvement may originate from the undocumented base system rather than the proposed components.

2. The ablation study is too coarse

Table 2 only adds CAC, OGR, and TRPL sequentially. It does not independently ablate the detail, geometry, and cross-band operators or the directional and contextual repair mechanisms. Action frequencies, repair success rates, and prediction changes before and after repair are also not reported.

3. Experimental coverage is limited

All results are reported on a single SpaceNet split, without repeated runs, standard deviations, statistical tests, per-city results, or cross-dataset evaluation. Given the complexity of the method, single-run results are insufficient to establish robustness.

Questions for authors

1.Were all Table 1 baselines rerun using exactly the same data split, preprocessing, input modalities, graph-construction procedure, and metric implementation? If not, which results were taken from prior publications?

2.Can the authors independently ablate the three CAC operators and the two OGR repair operators?

3.Can the authors report actual inference latency, peak GPU memory, and throughput in addition to dense TFLOPs?

Reviewer_fLUjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  • VERA-Road adds a region-wise observe–act–verify–repair loop to a frozen road-graph extractor consisting of a ViT encoder and road readout.
  • The first component, CAC, uses a LoRA-tuned Qwen2.5-VL-7B controller to inspect local, neighbouring, and global context alongside the current predictions. It selects the strengths of three residual feature operators: Haar detail, directional/dilated geometry, and multispectral cross-band processing.
  • The second component, OGR, sends regions with an unresolved score to another VLM query. The controller then chooses between rotation-averaged and multi-scale re-encoding.
  • The policy is trained through TRPL: supervised fine-tuning on searched, high-scoring actions, followed by a GRPO-style objective. The reward combines improvements in TOPO-F1 and APLS with an action-cost penalty.
Critical strengths
  • I like the idea of treating feature processing as a regional decision that can be checked and revised. Training the controller with graph-level rewards is a sensible way to address the lack of action annotations. The general approach could also be relevant to other structured-prediction tasks.
  • I also appreciate that the paper reports the substantial computational cost and discusses limitations. This makes it easier to assess the practical trade-off.
Critical weaknesses
  • The headline comparison is difficult to interpret because baseline comparability is unclear. The frozen Base already achieves 82.56 APLS, approximately eight points above the strongest external baseline in Table 1, despite having the lowest F1, at 80.73. Its recall, inferred from the reported F1 and precision, is approximately 73, which would also be the lowest. Different metrics can rank methods differently, so this is not necessarily an error. However, the size of the discrepancy needs explaining, particularly because missed connections can strongly affect APLS. Differences in graph post-processing or metric implementation are possible explanations. For context, SAM-Road reports F1 of 80.52 and APLS of 71.64 on SpaceNet. Its F1 is almost identical to the Base’s, while its APLS is roughly 11 points lower. The paper needs to establish whether these results were obtained under comparable conditions.
  • The experiments do not establish the need for a 7B VLM controller. The computational cost rises from 17.8 to 132 TFLOPs per image, with 8.39B additional parameters. Given this increase, comparisons with simpler policies are important. The paper does not evaluate always-on operators, random actions at a matched frequency, rules based on the existing uncertainty scores, or a small learned gate.
Questions for authors
  • What exactly is the Base model, including its encoder, readout, training data, and checkpoint? Were the Table 1 baselines rerun on your split using the same APLS implementation, or were their scores taken from published papers? Please include SAM-Road and SAM-Road++, and ideally evaluate on the standard 400×400, 1 m protocol.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.