Stereo Vision

About

Implementation of stereo vision in Python using image pairs from the Middlebury dataset.
The main idea is: given a left and right photo of the same scene, we compute a disparity map and then estimate a depth map.

I used OpenCV’s built-in stereo matchers:

• StereoBM (Block Matching)
• StereoSGBM (Semi-Global Block Matching)

---

What it does

• Takes a rectified stereo pair (left/right)
• Computes a disparity map
• Converts disparity to a depth map (approximate)
• Saves output images (and/or displays them)

---

Requirements

• Python 3.x
• OpenCV (opencv-python)
• NumPy

pip install opencv-python numpy

---

Dataset (Middlebury)

This project uses stereo pairs from the Middlebury dataset.

Download

You need to download the dataset separately (it’s not included in this repo).

• Dataset version: Middlebury 2021

---

How to run

python main.py

Methods (quick explanation)

StereoBM (Block Matching)

• Compares small windows (blocks) between the left and right image
• Faster and simpler
• Usually noisier / more “blocky” results

StereoSGBM (Semi-Global Block Matching)

• Uses a smoother optimization idea compared to BM
• Often produces cleaner disparity maps
• Slower but typically better quality

---

Depth map (from disparity)

Depth is inversely related to disparity. A common approximate relationship is:

depth ≈ (focal_length * baseline) / disparity

If calibration values aren’t available, the depth map is often “relative” (useful for visualization, but not exact meters).

---

Results

---

C++ implementation

A native OpenCV build lives under cpp/ (CMake + optional vcpkg manifest). It mirrors the Python pipeline: same calibration and stereo matchers, disparity → depth, visualization and optional export.

See cpp/README.md for configure, build, and run instructions.

---

Tech stack

• Python, OpenCV, NumPy
• C++20, OpenCV 4 (see cpp/)