Intersection-Flow-5k: Urban Traffic Detection Dataset

• Intersection-Flow-5k is a high-resolution urban intersection benchmark featuring over 6,900 images and 406,000 annotated objects across eight traffic categories.
• The dataset incorporates severe occlusion, extreme scale variation, and complex illumination to challenge detection models in realistic traffic scenarios.
• Benchmarking with FlowDet demonstrates improved accuracy and computational efficiency, making it a valuable resource for advancing real-time urban traffic monitoring.

The Intersection-Flow-5k dataset is a large-scale, high-resolution benchmark designed for traffic detection in urban intersection environments. Developed to evaluate model robustness under conditions characterized by dense traffic, severe occlusion, extreme scale variation, and complex illumination, Intersection-Flow-5k fills a gap in the domain of real-world intersection surveillance and algorithm benchmarking (Zhao et al., 27 Aug 2025). Its composition, annotation protocols, and usage scenarios facilitate the development and assessment of detection models suited to challenging urban transportation scenarios.

1. Dataset Structure and Composition

Intersection-Flow-5k comprises 6,928 static images with a resolution of $1920 \times 1080$ pixels, sourced from fixed, infrastructure-based cameras monitoring urban intersections. The dataset contains over 406,000 instance annotations across eight traffic-related categories:

• Vehicle
• Bus
• Bicycle
• Pedestrian
• Engine
• Truck
• Tricycle
• Obstacle

The dataset is organized into three splits for reproducible evaluation:

Split	Images	Total Objects
Training	5,483	327,825
Validation	722	38,868
Test	723	40,065

Table 1 in the source provides detailed breakdowns per class. Each object is labeled via bounding boxes and categorized according to these predefined classes. The annotation process specifically handles varying levels of occlusion (with up to 75% object occlusion recorded), and pays particular attention to capturing very small and large objects. The "obstacle" category encompasses diverse non-vehicle traffic elements relevant to intersection analysis.

2. Scene Characteristics and Annotation Protocol

Intersection-Flow-5k scenes are deliberately curated to mirror real-world complexities present in urban intersections. These include:

• Severe Occlusion: Many frames contain objects partially or mostly obscured by others, especially in crowded or congested states. Objects under heavy occlusion remain annotated to support rigorous detection evaluation.
• Extreme Scale Variations: Object sizes range from large, close-proximity vehicles to distant traffic participants barely 15×15 pixels in area. This demands algorithms capable of precise localization across orders of magnitude of apparent object scales.
• Perspective Distortion: Fixed cameras introduce perspective effects such as foreshortening and non-affine deformations, which are visually apparent in vehicles and other objects situated away from the principal axis.
• Illumination Challenges: The dataset includes varying ambient conditions, such as nighttime scenes with glare and sensor saturation from headlights, in addition to daylight occlusion and shadow cases.

Figure 1 of the source visually demonstrates key scenarios:

• (a) Sensor saturation during night-time glare,
• (b) Persistent occlusion of vehicles,
• (c) Detection of extremely small objects,
• (d) Baseline clear conditions.

This diversity ensures that any detection model evaluated on Intersection-Flow-5k faces a breadth of challenges encountered in practical surveillance.

3. Statistical Analysis

Detailed object counts highlight the dataset’s complexity:

Class	Train	Val	Test	Total
Vehicle	190,417	23,450	23,978	237,845
Bus	2,816	315	322	3,453
Bicycle	23,396	2,771	3,051	29,218
Pedestrian	19,005	2,250	2,459	23,714
Engine	642	74	80	796
Truck	7,749	881	916	9,546
Tricycle	1,610	225	200	2,035
Obstacle	82,190	8,902	9,059	100,151

Objects per image and class distribution reflect realistic, highly dynamic intersection environments. This granularity is critical for developing and validating detection and tracking systems.

4. Targeted Detection Challenges

Intersection-Flow-5k was constructed to explicitly facilitate research into:

• Detection under Heavy Occlusion: Labeling and testing with objects that are partially invisible, which is common around stationary buses, trucks, or clusters of pedestrians.
• Small Object Detection: Evaluation of models on objects with minimal pixel occupation, a known weak point for many current models.
• Dealing with Perspective Distortion: Requirements for architectural changes or data augmentations to preserve performance across the image.
• Illumination and Sensor Saturation: Robustness under real-world lighting, glare, reflections, and corresponding sensor limitations.

A central benchmarking goal is for models to generalize across the depicted scene diversity, rather than optimizing for only clear or canonical cases.

5. Benchmarking with FlowDet and Metrics

Intersection-Flow-5k supports standardized benchmarking for detection models, with particular emphasis given to the FlowDet detector (Zhao et al., 27 Aug 2025). FlowDet, evaluated on this dataset, achieves state-of-the-art results in both accuracy and computational efficiency:

• Accuracy Metrics:
  • AP (test set) = 58.6%, compared to RT-DETR’s 57.1% (improvement: +1.5%)
  • AP$_{50}$ (IoU threshold 0.5) = 82.3%, compared to 80.7% (+1.6%)
  • Small-object AP$_S$ = 34.2%, with baseline at 31.0%
• Computational Efficiency:
  • GFLOPs reduced by 63.2% (FlowDet: 50.0 GFLOPs, RT-DETR: 136.0 GFLOPs)
  • Inference speed of 136 FPS (versus 117 FPS for baseline)

These metrics are drawn from Table 2 and Figure 2 of the source, which illustrate the Pareto optimality of FlowDet for both speed and accuracy. Models are evaluated using accepted object detection metrics (AP, AP$_{50}$, class-specific and scale-based APs) as per standard object detection practice, and performance improvements are phrased as:

$AP^{test}_{FlowDet} = AP^{test}_{RT-DETR} + 1.5\%$

Thus, Intersection-Flow-5k plays an essential role as a performance benchmark in the context of modern real-time traffic detection systems.

6. Significance, Availability, and Role within the Research Landscape

The introduction of Intersection-Flow-5k addresses critical gaps in current intersection perception benchmarks. Its contributions include:

• Focus on intersection surveillance from infrastructure-based camera viewpoints.
• Representation of occluded, small, and variably illuminated objects required for robust detection.
• Availability for the community at https://github.com/AstronZh/Intersection-Flow-5K, with full split descriptions and annotation protocol.
• Application as a primary benchmark for advanced detection architectures such as FlowDet, enabling comparison both in accuracy and computational demand.

Relative to other datasets—such as A9 Intersection, IPS300+, and high-altitude UAV-based benchmarks—Intersection-Flow-5k is characterized by static, high-resolution footage at ground-level perspectives, systematic annotation of severe occlusion and small-scale objects, and its explicit use in challenging detector evaluation scenarios.

7. Applications and Research Directions

Intersection-Flow-5k underpins several active research areas:

• Real-time Intersection Monitoring: Enabling deployment of detection systems for traffic control, congestion monitoring, and safety analytics.
• Edge Deployment: The demonstrated performance and efficiency results from the FlowDet model suggest that, when evaluated on Intersection-Flow-5k, detectors can achieve real-time inference on resource-constrained platforms.
• Downstream Traffic Analysis: Its diversity and annotation fidelity make it suitable for augmenting tracking, trajectory prediction, or flow estimation research.
• Methodological Development: Intersection-Flow-5k provides a rigorous proving ground for architectural innovations (e.g., scale- and perspective-aware designs), hyperparameter tuning, and domain adaptation techniques pertinent to urban interoperability.

Models that succeed on Intersection-Flow-5k are plausibly robust to the most frequent and problematic sources of error in urban intersection applications, supporting both academic research and practical transportation deployments.