Semantic Segmentation

I have created a Semantic Segmentation model to segment potholes from an input video.

About the model

• I have trained a U-Net model from scratch using Tensorflow and Keras.
• I have used a custom Data Generator to feed the images and the masks as input to the U-Net model.
• The depth of the U-Net model can be customised. (recommended depth - 3 to 5 for (256,256,3) images).
• The maximum allowed depth of the U-Net model is determined by the dimensions of the input frames.

Structure of the U-Net Model

Run Locally

Following were the versions of the libraries I used

• Tensorflow - 2.10.1
• OpenCV (cv2) - 4.6.0
• Numpy - 1.23.5
• Moviepy - 1.0.3

Clone the project

  git clone https://github.com/lilNewbie/SemanticSegmentation.git

Go to the project directory

  cd SemanticSegmentation

Important! Set the paths to the input and output video files

Run segfile.py

  python segfile.py

Running the Train file (SemanticSegmentation_Pothole.ipynb)

• The U-Net model was trained in Google Colab.
• You will need to use your Kaggle public API to access the dataset via Colab

Sample Video

Input Video

input_video.mp4

Output Video

output_video.mp4

The difference in input and output clarity is due to resizing the input frames for prediction and later resizing it back to its actual dimensions

References

• Pothole Image segmentation dataset
• U-Net architecture