This repository explores fundamental techniques for estimating depth from stereo image pairs. It focuses on extracting disparity information from two grayscale images to generate accurate depth maps. The project implements both pixel-wise and window-based matching algorithms, utilizing different distance metrics and similarity measures.

This project includes the following depth estimation methods:

This method computes the disparity map by comparing individual pixel intensities between the left and right images. It calculates the cost for each possible disparity by measuring the pixel-wise distance between corresponding pixels, and then chooses the disparity with the minimum cost.

• Distance Metrics:
  • L1 Distance (Manhattan Distance): Calculates the absolute difference between pixel intensities.
  • L2 Distance (Euclidean Distance): Calculates the squared difference between pixel intensities.

This approach calculates disparities by comparing small image patches (windows) centered around each pixel. It improves the robustness to noise by considering a neighborhood of pixels instead of only individual pixels.

• Distance Metrics:
  • L1 Distance (Manhattan Distance): Calculates the absolute difference between pixels in the window.
  • L2 Distance (Euclidean Distance): Calculates the squared difference between pixels in the window.
• Similarity Measure:
  • Cosine Similarity: Measures the cosine of the angle between the vectors formed by pixel values within the corresponding windows.

The project consists of the following Python files:

• README.md: This file, providing an overview of the project.
• pixel_wise_matching.py: Implements pixel-wise matching with L1 and L2 distance metrics.
• window_based_matching.py: Implements window-based matching with L1 and L2 distance metrics.
• window_based_cosine_similarity.py: Implements window-based matching using cosine similarity as a matching metric.

To use this project:

1. Clone the repository:

2. Ensure you have the required libraries:

   • opencv-python
   • numpy Install these libraries using pip:

   pip install opencv-python numpy

3. Run the Python scripts:

   • To run the pixel-wise matching methods:

     python pixel_wise_matching.py

   • To run the window-based matching methods with L1 and L2 distance:

     python window_based_matching.py

   • To run the window-based matching method with cosine similarity:

     python window_based_cosine_similarity.py

4. View results:

   The generated disparity maps (both grayscale and colorized) will be saved in the respective result subdirectories:

   • result/pixel_wise/ for pixel-wise matching results.
   • result/window_based/ for window-based matching results.

The project uses the following image pairs:

• Tsukuba: Used for pixel-wise matching, located in assets/tsukuba/.
• Aloe: Used for window-based matching, located in assets/Aloe/.

The scripts will output the resulting disparity maps as grayscale images and colorized depth maps using the COLORMAP_JET color map.

• The scale variable controls the scaling of the calculated disparity before converting it to an 8-bit grayscale image.
• The kernel size is a crucial parameter that influences the performance of window-based matching.
• The disparity range should be adjusted based on the specific image pair being used.

• Implement additional matching algorithms (e.g. graph cuts, semi-global matching).
• Experiment with different window sizes and shapes for window-based methods.
• Apply post-processing techniques (e.g. median filtering) to reduce noise and improve the visual quality of depth maps.
• Implement a more robust cost aggregation method.

This project serves as a basic foundation for exploring and understanding stereo depth estimation. It can be extended to more sophisticated methods and used as a base for creating various applications that require depth information.