UAV-Crack

UAV-based pavement crack segmentation
  • Acquiring pavement distress images with UAVs faces unique challenges compared to ground-based methods due to differences in camera setups, flight parameters, and lighting conditions. These factors cause domain shifts that reduce the generalizability of segmentation models.
    • 图片1
  • A dedicated dataset, UAV-CrackX, was created with 1500 pixel-wise annotated UAV images to support model development.The dataset includes three subsets—UAV-CrackX4, X8, and X16—with 500 images each at 4×, 8×, and 16× zoom levels.
    • 图片2
  • Original high-resolution images (2688 × 1512) were split into 16 patches (672 × 378 pixels) for efficient processing. Reference baseline model: https://github.com/open-mmlab/mmsegmentation/tree/main
Active
Click to enter and fill out the application form Get Dataset
Click to manage your methods or submit results Submit Task

Submission Guidelines

UAV-Crack Benchmark Evaluation - Data Format & Submission Guidelines

UAV-Crack Benchmark Evaluation

1. Evaluation Data Format

After unzipping the dataset you received, you will find the following folder structure:

  • gtFine folder: Contains binary mask images output by your model.
    Pixel value 0 represents normal road surface, 1 represents cracks.
  • train subfolder (inside gtFine): Includes binary images at 3 scales: X4, X8, X16, all with resolution 672×378.
  • leftimg8bit folder: Contains input images (original road images).
  • train subfolder (inside leftimg8bit): Input images at 3 scales: X4, X8, X16, resolution 672×378.
  • val subfolder (inside leftimg8bit): Contains 300 validation images for testing, also 672×378.
  • showgt.py: A visualization script to display binary masks more intuitively (not used in evaluation).

Example of Model Input Image (.jpg):

Input Image Example

Example of Model Output Binary Mask (.png):

Output Binary Mask Example

Example of Visualized Binary Mask using showgt.py:

Visualized Mask Example

Benchmark Metrics Introduction:

  • "mIoU": Mean Intersection over Union
  • "Crack F1": Harmonic mean of Precision and Recall
  • "Crack IoU": Intersection over Union for crack class
  • "Crack Precision": Precision for crack detection
  • "Crack Recall": Recall for crack detection
  • "aAcc": Overall Accuracy (Average Accuracy)

2. Evaluation Result Submission Format

We have selected 300 images from the leftimg8bit/val/ folder for evaluation. You are required to submit your model’s predicted binary crack masks for these images only.

Submission Requirements:

  • 1. The resolution of your predicted binary images must be the same as the input, which is 672×378.
  • 2. The file names of your predicted images must exactly match the input image names (but with .png extension).
  • 3. Note: Input images are in .jpg format, but your output predictions must be in .png format.

How to Submit:

  • Please compress all 300 predicted binary mask images (.png) into a single zip file.
  • The zip file should be named: result.zip

Important: Do not include any extra files or images in the submission zip. Only include the 300 required predicted mask images. This ensures correct evaluation.

Method Leaderboard

0 Methods 0 Metrics
No methods available for this task yet.