Potholes

The simulated potholes were made to mimick potholes that would be present at the Intelligent Ground Vehicles Competition 2022. The rules specified a "Solid white circle with 2' diameter".

Generating Synthetic Data

Using this technique of data creation was highly efficient for two reasons. Firstly, I was able to reuse a dataset of 1.6k images that were initially used to train a U-Net model to detect lane lines. By using the same dataset for potholes, I can save time on collecting data. Secondly, the YOLOv4 model requires a label around each object in the format of <x> <y> <width> <height>. I can easily create a label for each pothole while generating them, saving time on labelling.

• In the data directory, add folders containing the input images. The script will then iterate through each of the images.

• For each image, a random number of potholes are generated (1, MAX_POTHOLES).

• The script finds a rectangular patch in the image, where the average color is grey (rgb value range from 15 to 180).

  • Only specific parts of the image are iterated through: Between 20%-80% on the horizontal axis and between the 70%-90% on the vertical axis.
  • These numbers were picked to guess where asphalt in a lane image is more likely to appear.
  • It also makes sure that potholes are not overlapping.

• After finding a suitable patch, an ellipse is generated with a grey color fill.

  • The greyness depends on the rgb values in the patch (Darker asphalt will have darker potholes).

• The location of the potholes in YOLOv4 format is then appended to a .txt containing all the potholes for that image.

  • The labelled image is stored in processed/data/.
  • The label is stored in processed/label/.

Example Images

Randomely generated potholes in Toronto streets.

Potholes with metaball noise to act as shadows.

Potholes with fine salt-pepper noise.

Splitting the data into train/test/val

(train, test, val) = (70%, 10%, 20%)

• Currently, the script simply takes data from dataset/data/ and dataset/label/ and moves it to dataset/train/, dataset/test/, or dataset/val/.
  • The images are chosen at random and therefore the images are shuffled.