A Road-Conditioned Traffic Movie Prediction Network with Spatiotemporal and Structure-Consistent Learning

Armstrong Aboah; Blessing Agyei Kyem; Joshua Kofi Asamoah

arxiv: 2605.27884 · v1 · pith:XRYWKAPI · submitted 2026-05-27 · cs.CV

A Road-Conditioned Traffic Movie Prediction Network with Spatiotemporal and Structure-Consistent Learning

Joshua Kofi Asamoah , Blessing Agyei Kyem , Armstrong Aboah This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-29 13:51 UTCgrok-4.3pith:XRYWKAPI record.json open to challenge →

classification cs.CV

keywords traffic forecastingspatiotemporal networksroad conditioningstructure consistencycross-city generalizationtraffic predictionurban mobility

0 comments

The pith

RCSNet uses static road maps and structure-consistent learning to improve traffic movie predictions in both same-city and cross-city settings.

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

This paper proposes RCSNet to address limitations in existing traffic forecasting methods that treat predictions as image reconstruction without sufficient road constraints. It claims that conditioning on static road maps and enforcing structure consistency allows the model to generate future traffic maps that are more accurate and aligned with road topology and direction. The approach is shown to reduce error metrics in forecasting on multiple cities and to generalize better to unseen cities. A sympathetic reader would care because accurate traffic predictions are key for congestion management and intelligent transportation systems.

Core claim

RCSNet reformulates traffic movie prediction as topology-guided future-state generation. It extracts road-aware representations from static road maps, models multi-horizon traffic dynamics from historical observations, aligns directional traffic features with local road structure, and progressively generates future traffic maps. A structure-consistent learning objective encourages predictions to remain accurate, road-aligned, and spatially stable. Experiments show improvements in forecasting accuracy and structural consistency across cities.

What carries the argument

The RCSNet architecture, which conditions spatiotemporal predictions on static road maps and uses a structure-consistent objective to enforce road alignment in generated traffic maps.

If this is right

Reduces average MAE by 11.5%, MSE by 10.0%, and RMSE by 5.1% in same-city forecasting on Berlin, Antwerp, and Moscow.
Reduces RMSE by 10.6% and 10.5% in cross-city testing on Chicago and Bangkok without fine-tuning.
Produces predictions with improved structural consistency and temporal stability.
Offers computational efficiency advantages in addition to accuracy gains.

Where Pith is reading between the lines

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

The method may enable better integration with route guidance systems by ensuring predictions respect travel directions.
Extending the approach to dynamic road conditions like construction or accidents could further enhance real-world applicability.
Similar conditioning techniques might apply to other spatiotemporal prediction tasks such as crowd flow or weather patterns over networks.
Longer prediction horizons might benefit more from the topology guidance if the structure consistency holds over time.

Load-bearing premise

Static road maps supply sufficient topology and directional information to guide multi-horizon traffic dynamics, and the structure-consistent objective enforces road alignment without introducing new biases or reducing accuracy on non-road features.

What would settle it

A test where RCSNet is applied to a city with significantly different road structures or traffic patterns and shows no reduction in errors or even increased structural misalignment compared to baselines would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.27884 by Armstrong Aboah, Blessing Agyei Kyem, Joshua Kofi Asamoah.

**Figure 1.** Figure 1: Overall architecture of the proposed RCSNet framework. Historical traffic frames and the static road map are processed through separate temporal and topology-aware branches, fused through direction-aware road-traffic interaction, and decoded progressively to generate future traffic maps under a structure-consistent learning objective. 3.3. Proposed Architecture Building on the input representation, the pro… view at source ↗

**Figure 2.** Figure 2: Detailed architecture of the topology-aware road representation module. The static road map is transformed into multiple structural prior channels, including occupancy, centerline, edge magnitude, orientation, connectivity, and intersection tendency, before being encoded through multi-scale spatial branches to produce topology-aware road features. of traffic channels, 𝑇𝑖𝑛 is the number of observed frames,… view at source ↗

**Figure 3.** Figure 3: Detailed architecture of the multi-horizon temporal traffic encoder. Historical traffic sequences are processed through parallel temporal convolution branches with different kernel sizes and dilation rates to capture short-term, intermediate-term, and long-term traffic dynamics before feature fusion. where [⋅; ⋅; ⋅] denotes concatenation along the channel dimension, 𝜓𝑡 (⋅) denotes the fusion projection, a… view at source ↗

**Figure 4.** Figure 4: Detailed architecture of the direction-aware road-traffic fusion module. The topology-aware road feature guides channel, spatial, and direction-aware modulation of the traffic representation, enabling the fused feature to reflect both dynamic traffic evolution and local road structure. The road-guided traffic feature, projected road feature, and direction gate are then concatenated: 𝐅 𝑐𝑎𝑡 = [ 𝐅̃𝑡𝑒𝑚𝑝 𝑡 ; 𝐅̃… view at source ↗

**Figure 5.** Figure 5: Qualitative comparison of traffic forecasting results in Berlin at 8:15am, 1:55pm, and 7:30pm. RCSNet preserves the main road-aligned traffic structure more closely than the baseline methods, with higher SSIM, lower MAE, and non-zero cell counts closer to the ground truth. road map through its topology-aware road representation and direction-aware fusion modules. As shown in [PITH_FULL_IMAGE:figures/full_… view at source ↗

**Figure 6.** Figure 6: Qualitative comparison of traffic forecasting results in Antwerp at morning, afternoon, and evening periods. RCSNet maintains compact traffic activations that closely match the ground truth, while several baselines produce broader residual errors and less road-aligned predictions. On Chicago, RCSNet obtains an MAE of 0.1312, MSE of 0.4986, and RMSE of 0.7061. Compared with the closest competing baseline, M… view at source ↗

**Figure 8.** Figure 8: Cross-city qualitative forecasting results on Chicago. The first three columns show morning, afternoon, and evening predictions, while the final column shows the aggregated absolute error heatmap across the selected times. 5.5. Forecast Horizon Analysis To further evaluate the temporal stability of the proposed model, forecast horizon analysis is conducted over five future horizons, namely 𝑡 + 5, 𝑡 + 15, … view at source ↗

**Figure 10.** Figure 10: Forecast horizon SSIM heatmap across Berlin, Antwerp, and Moscow. RCSNet maintains the strongest performance across short- and long-term horizons, showing slower degradation from 𝑡 + 5 to 𝑡 + 60 minutes compared with the baseline methods. 5.6. Road-Structure Consistency Analysis To further examine whether the predicted traffic maps respect the physical road network, a road-structure consistency analysis… view at source ↗

**Figure 11.** Figure 11: Road-structure consistency comparison across Berlin, Antwerp, and Moscow. RCSNet achieves lower roadregion error, fewer off-road activations, and higher road coverage recall, indicating stronger alignment between predicted traffic maps and the underlying road network. RCSNet and each baseline method, as shown in [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗

read the original abstract

City-wide traffic forecasting is important for congestion management, route guidance, and intelligent transportation systems, but accurate prediction remains challenging when future traffic must be generated as spatial maps over an entire urban network. Existing traffic movie prediction methods have improved frame-level accuracy, yet many still treat forecasting mainly as image reconstruction. This can produce traffic maps that are numerically close to the ground truth but weakly constrained by road layout, connectivity, travel direction, and congestion propagation, especially in cross-city settings where both traffic behavior and road structure change. To address this limitation, this study proposes RCSNet, a road-conditioned spatiotemporal network that reformulates traffic movie prediction as topology-guided future-state generation. RCSNet extracts road-aware representations from static road maps, models multi-horizon traffic dynamics from historical observations, aligns directional traffic features with local road structure, and progressively generates future traffic maps for improved temporal consistency. A structure-consistent learning objective further encourages predictions to remain accurate, road-aligned, and spatially stable. Experiments across multiple cities show that RCSNet improves both forecasting accuracy and structural consistency. In same-city forecasting on Berlin, Antwerp, and Moscow, RCSNet reduces average MAE, MSE, and RMSE by 11.5%, 10.0%, and 5.1%, respectively, compared with the closest baseline. In cross-city testing on unseen Chicago and Bangkok, it reduces RMSE by 10.6% and 10.5% without target-city fine-tuning. Additional horizon-wise, road-structure, explainability, statistical, and efficiency analyses show that RCSNet produces more accurate, transferable, road-aligned, and computationally efficient traffic forecasts.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RCSNet layers road conditioning and a structure loss onto spatiotemporal predictors and reports cross-city transfer gains, but the abstract supplies no baselines or ablations to show the new pieces are responsible.

read the letter

RCSNet conditions traffic map prediction on static road maps, adds directional alignment and a structure-consistent loss, then claims better accuracy and transfer. The headline result is the cross-city RMSE drop of roughly 10.5% on Chicago and Bangkok with no fine-tuning.

The combination of road-aware extraction, progressive generation, and the structure objective is presented as a new formulation on top of existing video prediction backbones. That is the concrete addition.

The work is useful for readers who need transferable urban traffic forecasts; the same-city gains on Berlin, Antwerp, and Moscow and the no-retraining cross-city numbers are the parts that could matter for intelligent transportation applications.

The soft spots sit in the evidence. The abstract states metric improvements but lists neither the baselines, the data splits, ablation results, nor any statistical tests. Without those, it is impossible to tell whether the reported 11.5% MAE reduction comes from the road conditioning or from other unstated changes. The stress-test concern also lands: static maps encode fixed topology and direction, yet multi-horizon traffic is also driven by time-varying signals, incidents, and demand shifts that the abstract does not model. If those factors dominate, the claimed mechanism is not yet shown to explain the gains.

The paper is aimed at the traffic forecasting subgroup of computer vision and transportation research. A reader working on cross-city generalization might extract value from the transfer experiments once the full methods and controls are visible.

I would send it to peer review so the experiments can be checked, but it is not yet a clear accept on the information given.

Referee Report

2 major / 2 minor

Summary. The paper proposes RCSNet, a road-conditioned spatiotemporal network for city-wide traffic movie prediction. It reformulates the task as topology-guided future-state generation by extracting road-aware representations from static road maps, modeling multi-horizon dynamics, aligning directional features with road structure, and applying a structure-consistent learning objective for accuracy, alignment, and spatial stability. Experiments on Berlin, Antwerp, Moscow, Chicago, and Bangkok report average reductions of 11.5% MAE, 10.0% MSE, and 5.1% RMSE in same-city forecasting versus the closest baseline, plus 10.6% and 10.5% RMSE reductions in cross-city transfer without fine-tuning, along with horizon-wise, structural, explainability, statistical, and efficiency analyses.

Significance. If the empirical gains hold under rigorous controls, the work could advance traffic forecasting by demonstrating that static topology and directional constraints improve both accuracy and cross-city transferability over pure image-reconstruction approaches. The structure-consistent objective and road-aware representations address a recognized limitation in existing spatiotemporal models.

major comments (2)

[Experiments and Ablation Studies] The central claim that road conditioning from static maps drives the reported MAE/MSE/RMSE reductions (both same-city and cross-city) requires evidence that topology and directionality are primary constraints on multi-horizon dynamics rather than time-varying signals. The manuscript should include ablations that isolate the road-map input and structure-consistent objective (e.g., comparing against a non-road-conditioned spatiotemporal baseline) to rule out that gains arise from other unstated factors.
[Cross-City Evaluation] Cross-city results on unseen Chicago and Bangkok are presented as evidence of transferability without target-city fine-tuning, yet the paper must demonstrate that the static road maps of the source cities (Berlin/Antwerp/Moscow) generalize to the target topologies; otherwise the 10.6%/10.5% RMSE reductions cannot be attributed to the proposed mechanism.

minor comments (2)

[Abstract and §4] The abstract and results sections should explicitly list the baselines, data splits, and statistical significance tests used to compute the percentage improvements.
[Method] Notation for the structure-consistent loss and road-aware feature extraction should be defined with equations before the experimental claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and the recommendation for major revision. We address each major comment below and outline the changes we will make to the manuscript.

read point-by-point responses

Referee: [Experiments and Ablation Studies] The central claim that road conditioning from static maps drives the reported MAE/MSE/RMSE reductions (both same-city and cross-city) requires evidence that topology and directionality are primary constraints on multi-horizon dynamics rather than time-varying signals. The manuscript should include ablations that isolate the road-map input and structure-consistent objective (e.g., comparing against a non-road-conditioned spatiotemporal baseline) to rule out that gains arise from other unstated factors.

Authors: We agree that explicit ablations isolating the road-map input and the structure-consistent objective would provide stronger evidence for our claims. Although the current experiments compare RCSNet against several spatiotemporal baselines that lack road conditioning, we will add dedicated ablation studies in the revised manuscript. These will include variants of RCSNet without the road-map encoder and without the structure-consistent loss term, allowing direct quantification of their contributions to the observed performance improvements. revision: yes
Referee: [Cross-City Evaluation] Cross-city results on unseen Chicago and Bangkok are presented as evidence of transferability without target-city fine-tuning, yet the paper must demonstrate that the static road maps of the source cities (Berlin/Antwerp/Moscow) generalize to the target topologies; otherwise the 10.6%/10.5% RMSE reductions cannot be attributed to the proposed mechanism.

Authors: We believe there may be a misunderstanding regarding the cross-city setup. In our experiments, the model is trained exclusively on traffic data from the source cities but, during inference on the target cities, we provide the target cities' static road maps as input to the road-conditioning module. This allows us to test whether the spatiotemporal dynamics learned from source cities can be applied to novel road topologies without retraining. We will revise the manuscript to explicitly describe this procedure and add a discussion on how the road-aware representations enable generalization across different topologies. If additional experiments are required, we can explore training without any road map input for cross-city as well. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ML model with experimental validation

full rationale

The paper introduces RCSNet as a neural architecture for road-conditioned traffic movie prediction and reports empirical MAE/MSE/RMSE reductions on held-out city data (Berlin, Antwerp, Moscow, Chicago, Bangkok). No equations, derivations, or self-citations appear that reduce any claimed prediction or uniqueness result to fitted inputs by construction. The structure-consistent objective and road-map conditioning are design choices whose performance is measured experimentally rather than asserted by definition or prior self-work. This matches the most common honest finding for applied CV papers: self-contained empirical results with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Only the abstract is available; the ledger is therefore populated at the level of standard deep-learning assumptions rather than paper-specific details.

free parameters (1)

network weights and hyperparameters
Typical for any neural network; fitted during training on traffic data.

axioms (1)

domain assumption Standard assumptions of supervised deep learning on spatiotemporal image sequences hold for traffic maps.
The model relies on typical neural network training and evaluation practices.

pith-pipeline@v0.9.1-grok · 5835 in / 1277 out tokens · 35059 ms · 2026-06-29T13:51:58.211715+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[1]

Journal of Transportation Engineering, Part B: Pavements 152, 04026010

Pavecap: A multimodal framework for comprehensive pavement condition assessment withdense captioning and pci estimation. Journal of Transportation Engineering, Part B: Pavements 152, 04026010. Ahn, J., Ko, E., Kim, E.Y., 2016. Highway traffic flow prediction using support vector regression and bayesian classifier, in: 2016 International conferenceonbigdat...

work page arXiv 2016
[2]

arXiv preprint arXiv:2503.21158

Integrating travel behavior forecasting and generative modeling for predicting future urban mobility and spatial transformations. arXiv preprint arXiv:2503.21158 . Diao, Z., Wang, X., Zhang, D., Liu, Y., Xie, K., He, S., 2019. Dynamic spatial-temporalgraphconvolutionalneuralnetworksfortrafficforecast- ing, in: Proceedings of the AAAI conference on artific...

work page arXiv 2019
[3]

A graph-based u-net model for predicting traffic in unseen cities, in: 2022 International Joint Conference on Neural Networks (IJCNN), IEEE. pp. 1–8. Ho,S.L.,Xie,M.,1998. Theuseofarimamodelsforreliabilityforecasting and analysis. Computers & industrial engineering 35, 213–216. Hong, W.C., Pai, P.F., Yang, S.L., Theng, R., 2006. Highway traffic fore- casti...

work page arXiv 2022
[4]

Decoupled dynamic spatial-temporal graph neural network for traffic forecasting,

Decoupled dynamic spatial-temporal graph neural network for traffic forecasting. arXiv preprint arXiv:2206.09112 . Sharma,S.,Majumdar,A.,Elvira,V.,Chouzenoux,E.,2020. Blindkalman filtering for short-term load forecasting. IEEE Transactions on Power Systems 35, 4916–4919. Shoman, M., Aboah, A., Daud, A., Adu-Gyamfi, Y., 2024. Graph con- volutional gated re...

work page arXiv 2020

[1] [1]

Journal of Transportation Engineering, Part B: Pavements 152, 04026010

Pavecap: A multimodal framework for comprehensive pavement condition assessment withdense captioning and pci estimation. Journal of Transportation Engineering, Part B: Pavements 152, 04026010. Ahn, J., Ko, E., Kim, E.Y., 2016. Highway traffic flow prediction using support vector regression and bayesian classifier, in: 2016 International conferenceonbigdat...

work page arXiv 2016

[2] [2]

arXiv preprint arXiv:2503.21158

Integrating travel behavior forecasting and generative modeling for predicting future urban mobility and spatial transformations. arXiv preprint arXiv:2503.21158 . Diao, Z., Wang, X., Zhang, D., Liu, Y., Xie, K., He, S., 2019. Dynamic spatial-temporalgraphconvolutionalneuralnetworksfortrafficforecast- ing, in: Proceedings of the AAAI conference on artific...

work page arXiv 2019

[3] [3]

A graph-based u-net model for predicting traffic in unseen cities, in: 2022 International Joint Conference on Neural Networks (IJCNN), IEEE. pp. 1–8. Ho,S.L.,Xie,M.,1998. Theuseofarimamodelsforreliabilityforecasting and analysis. Computers & industrial engineering 35, 213–216. Hong, W.C., Pai, P.F., Yang, S.L., Theng, R., 2006. Highway traffic fore- casti...

work page arXiv 2022

[4] [4]

Decoupled dynamic spatial-temporal graph neural network for traffic forecasting,

Decoupled dynamic spatial-temporal graph neural network for traffic forecasting. arXiv preprint arXiv:2206.09112 . Sharma,S.,Majumdar,A.,Elvira,V.,Chouzenoux,E.,2020. Blindkalman filtering for short-term load forecasting. IEEE Transactions on Power Systems 35, 4916–4919. Shoman, M., Aboah, A., Daud, A., Adu-Gyamfi, Y., 2024. Graph con- volutional gated re...

work page arXiv 2020