Calibrate Your Own Webcam

Calibrating a camera means recovering the numbers that turn 3D points into pixels: the focal length, the principal point, the lens distortion, and how the camera is tilted. In this take-home midterm you calibrate your own laptop webcam three ways. First you use a checkerboard and take its method apart step by step. Next you use GeoCalib, a neural network that calibrates from one ordinary photo. You study what the network predicts before you rely on it. Last, you calibrate from a video of yourself walking, with the single-view geometry that NVIDIA’s AutoMagicCalib uses on camera networks. The checkerboard result is your ground truth for the other two.

​

Grading

​

Background reading

• From World Points to Pixels, sections 39.7.1 to 39.7.4 on checkerboard calibration.
• Single View Metrology, on vanishing points and the horizon line.
• For the last question: sections 44.2 and 44.3 of the MIT book’s chapter on multiview geometry and structure from motion.

​

How the tasks work

Each task asks for one function or a few lines. Most tasks are followed by a check cell. It prints passed when your code is right and explains what is wrong otherwise. Later parts use your earlier functions, so do not move on until a check passes. Some results depend on your camera and your room, not only on your code. Those checks print a warning instead of failing. The questions ask you to explain what you did and observed, in your own words, with your own numbers. Until your own captures are in place, the code runs on a sample captured on another laptop, so you can start the tasks right away. The results you hand in must come from your own webcam. Any part answered with the sample data is penalized. This is an individual exam. You may consult the course material, the textbooks and the documentation of the libraries the code uses, but the code and the written answers must be your own.

​

What you will need

• Your laptop’s built-in webcam. Use the same camera, at the same resolution, for every capture.
• A checkerboard, printed flat or shown full screen on a phone or tablet. Count its inner corners (a board with 11 by 8 squares has 10 by 7) and measure the side of one square in millimeters.
• A phone inclinometer app, to read the angle of your laptop lid. Only changes between readings are used, so the app’s zero does not matter.
• The capture script, which runs on your laptop with Python, opencv-python and numpy. Colab cannot open a local camera.
• A Google account for Colab and Google Drive, and a GPU runtime: in Colab, open Runtime, select Change runtime type, and choose a T4 GPU.

​

What to capture

Run python capture.py list to find your color webcam, then python capture.py capture --device <index> --pattern <columns> <rows> --square-mm <size> with your board’s inner-corner count and square size. In the preview window, c saves a checkerboard view, s saves an ordinary scene, v starts and stops a video, and q quits. The first cells of the midterm describe each capture in detail.

• About 20 checkerboard views, covering the whole image, with the board tilted in different directions.
• About 10 ordinary scenes with no checkerboard, each a different view, plus four scenes of one view at four lid angles, with the phone reading for each.
• A 30 to 60 second video of yourself walking across the room, with the laptop still, your whole body in view, and the phone reading on the lid noted.

Upload the whole captures folder to My Drive/cv-midterm/captures in your Google Drive. The setup cell mounts your Drive and reads it from there, so your data survives the end of a Colab session and never goes into your repository.

​

What you hand in

• The notebook, with every cell output saved.
• In the notebook’s last cell, a view-only link to your captures folder in Google Drive.

​

Assignment notebook

​

Submitting

Commit the notebook to your private clone of the class repo under assignments/midterm. Save every cell output before committing. A notebook without outputs is graded as though its cells never ran. Do not commit your captures. Submit the complete GitHub URL for that directory to Canvas or Brightspace. Do not submit the repository root. The assignment submission guide explains the remaining requirements, including how to set up your clone, add the TAs as collaborators, and create the README required in each assignment folder.

---