Abstract
The TLab team secured 2nd place in the NeurIPS 2020 Traffic4cast Challenge by treating citywide traffic forecasting as a high-resolution "movie" prediction task. Their winning approach leverages HR-NET (High-Resolution Network) combined with a novel geo-embedding technique that allows the model to learn the unique, inherent traffic signatures of specific geographic locations.
Introduction
Predicting traffic flow and speed across an entire city is a daunting spatio-temporal challenge. Traditional methods often struggle to scale or fail to capture the complex relationship between time (rush hours, holidays) and space (residential streets vs. major highways). Most computer vision models assume "translation invariance"—meaning a feature like a "traffic jam" should be recognized the same way regardless of where it appears in the image. However, in urban planning, location matters immensely; a pixel representing a cul-de-sac behaves fundamentally differently than one representing a four-lane bridge.
The authors sought to bridge this gap by creating a model that understands both the visual patterns of traffic and the unique identity of every "pixel" in the city. This study utilizes the Traffic4cast 2020 Challenge dataset, which provides industrial-scale, real-world traffic data for Berlin, Istanbul, and Moscow. Each city is mapped to a 436 × 495 grid (100m × 100m per pixel), with data aggregated into 5-minute bins over a full year.
Methodology
The authors moved away from the standard U-Net architecture, opting instead for HR-NET as their backbone. Unlike architectures that compress and then recover resolution, HR-NET maintains high-resolution representations throughout the entire process, which is critical for preserving fine-grained traffic details.
Key technical contributions include:
- Geo-Embedding: By assigning a unique ID to every pixel in the city grid and using a learnable embedding layer, the model learns "personalized" attributes for every region. This breaks translation invariance where spatial context is paramount.
- Feature Engineering: The problem is framed as supervised learning, using 12 frames of historical data (one hour) to predict the next 6 frames. Hand-crafted channels for periodic statistics, weekdays, and holiday status provide strong temporal priors.
- Optimization: The team utilized ELU activation functions, the LAMB optimizer, and learning rate warm-up to stabilize training on high-dimensional spatio-temporal data.
Results and Discussion
The TLab solution was highly competitive, taking 2nd place globally. HR-NET-W48 achieved a Mean Squared Error (MSE) of 1.2919e-3 on the Berlin dataset, consistently outperforming U-Net variants.
Ablation studies confirmed the efficacy of the proposed components. Adding Periodic Features dropped the Berlin MSE from 1.3040e-3 to 1.2980e-3. Furthermore, the use of geo-embeddings and the LAMB optimizer provided measurable boosts in validation performance.
Key Insights
- Identity Matters in Space: The success of geo-embeddings suggests that the "meaning" of a traffic pattern is coordinate-dependent.
- Periodicity as a Prior: Providing "daily average statistics" allows the model to leverage cyclical city rhythms without relearning them from scratch.
- High Resolution is King: Maintaining high-resolution feature maps via HR-NET is superior for capturing localized traffic bottlenecks that are often lost during downsampling.
Conclusion
While the model excels at capturing periodic patterns, it remains vulnerable to external shocks like accidents or extreme weather. Future research will focus on integrating these external influences to enhance robustness against anomalies that deviate from the "average" city rhythm.
Summary updated