Welcome to the official repository for the Intersection-Flow-5K dataset, introduced in our paper:

FlowDet: Overcoming Perspective and Scale Challenges in Real-Time End-to-End Traffic Detection

Yuhang Zhao, Zixing Wang

PRCV, 2025

This dataset is specifically designed to address the unique challenges of real-world, infrastructure-based traffic monitoring. It features high-density scenes, extreme scale variations, and severe, persistent occlusions, providing a challenging benchmark for modern object detectors.The Intersection-Flow-5k dataset is available at dataset.

Existing object detection benchmarks often fall short of capturing the complexities of fixed-camera traffic surveillance. Intersection-Flow-5K was created to fill this gap, offering a unique set of challenges:

• Extreme Scale Variation: Objects range from distant vehicles appearing as small as 15x15 pixels to large trucks occupying over 800x600 pixels within a single frame.
• High Object Density & Severe Occlusion: The dataset includes rush-hour scenes with numerous overlapping vehicles, leading to persistent and severe occlusions (up to 75% annotated).
• Diverse Environmental Conditions: Data was collected from 7 distinct urban intersections, covering various times of day (daylight, nighttime with glare) and weather conditions (clear, overcast, rainy).
• Comprehensive Annotations: Meticulously annotated with high-quality bounding boxes for crucial traffic participants.

This benchmark is ideal for researchers working on robust object detection, small object detection, detection in crowded scenes, and real-time intelligent transportation systems (ITS).

• Task Type: Object Detection

• Total Images: 6,928 high-resolution (1920x1080) images

  • Training Set: 5,483 images (80%)
  • Validation Set: 722 images (10%)
  • Test Set: 723 images (10%)

• Total Annotations: Over 95,000 bounding boxes

• Number of Categories: 8

• Category List (classes.txt):

  vehicle
  bus
  bicycle
  pedestrian
  engine
  truck
  tricycle
  obstacle

The dataset is organized as follows:

Intersection-Flow-5K/
├── images/                   # Original high-resolution images
│   ├── train/                # 5,483 images
│   ├── val/                  # 722 images
│   └── test/                 # 723 images
│
├── labels/                   # Annotations in YOLO .txt format
│   ├── train/
│   ├── val/
│   └── test/
│
├── annotations/              # Annotations in PASCAL VOC .xml format
│   ├── train/
│   ├── val/
│   └── test/
│
├── test_coco.json            # Annotations for the test set in COCO .json format
│
├── intersection.yaml         # Dataset configuration file for YOLO
├── classes.txt               # List of class names
│
├── convert_coco.py           # Example script for coco format conversion (optional)
│
└── README.md            
│
└── README_zh.md  

To maximize compatibility with various detection frameworks, we provide annotations in three standard formats: YOLO, PASCAL VOC, and COCO (for the test set).

Located in the labels/ directory. Each image has a corresponding .txt file where each line represents an object.

• Format: <class_id> <x_center> <y_center> <width> <height> (all values are normalized to [0, 1]).
• Example (image_001.txt for a 1920x1080 image):

  0 0.5416 0.6111 0.1041 0.1851  # A 'vehicle' object
  5 0.2343 0.7870 0.1562 0.2222  # A 'truck' object

Located in the annotations/ directory. Each image has a corresponding .xml file containing object bounding boxes in absolute pixel coordinates.

• Format: Bounding boxes are defined by <xmin>, <ymin>, <xmax>, <ymax>.
• Example (image_001.xml):

  <annotation>
      ...
      <size>
          <width>1920</width>
          <height>1080</height>
      </size>
      <object>
          <name>vehicle</name>
          <bndbox>
              <xmin>940</xmin>
              <ymin>560</ymin>
              <xmax>1140</xmax>
              <ymax>760</ymax>
          </bndbox>
      </object>
      ...
  </annotation>

We provide test_coco.json for easy evaluation of the test set using standard COCO evaluation scripts. The convert_coco.py script can be used as a reference to convert YOLO/VOC annotations to the COCO format if needed for your custom splits.

This dataset is ready to use with modern YOLO frameworks. Just point your training command to the provided intersection.yaml file.

1. Modify intersection.yaml: Ensure the paths in the YAML file are correct relative to your project's root directory.

   # intersection.yaml
   path: /path/to/Intersection-Flow-5K  # IMPORTANT: Update this path
   train: images/train
   val: images/val
   test: images/test
   
   # Classes
   nc: 8
   names: ['vehicle', 'bus', 'bicycle', 'pedestrian', 'engine', 'truck', 'tricycle', 'obstacle']  

2. Start Training:

   # Example for YOLOv8
   yolo detect train data=intersection.yaml model=yolov8l.pt epochs=200 imgsz=640 ...

These frameworks typically support PASCAL VOC or COCO formats.

• For VOC Format: Point your data configuration to use the images/ and annotations/ directories. You will need to provide a file list for the train, val, and test splits.
• For COCO Format: Use the provided test_coco.json for evaluation. For training/validation, you can use the convert_coco.py script to generate train.json and val.json from the VOC/YOLO annotations.

If you use the Intersection-Flow-5K dataset or find our work on FlowDet beneficial to your research, please cite our paper:

@inproceedings{zhao2025flowdet,
  title     = {FlowDet: Overcoming Perspective and Scale Challenges in Real-Time End-to-End Traffic Detection},
  author    = {Yuhang Zhao and Zixing Wang},
  booktitle = {PRCV},
  year      = {2025}
}

This dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. You are free to share and adapt the material for non-commercial purposes, provided you give appropriate credit.