Overview

Implementing camera calibration and structure from motion (SfM) methods to produce a reconstruction of a scene. Using the Direct Linear Transform (DLT) method to compute the scene geometry from correspondences.

Setup

We recommend to use a virtual environment to avoid cluttering your Python installation. To set this up using the Python venv module on Linux just use

cd code
python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
In case you want to use a different setup, just install the dependencies in requirements.txt manually.

Data

The data folder contains all the data used for this project and is structured the following way:

`--data
   `--calibration
      |--image_size.txt
      |--points2d.txt
      |--points3d.txt
   `--images
   `--keypoints
   `--matches

Calibration

The files contain informations about the 3D points of a scene and the 2D points of the scene seen from the camera. image_size only has two values for width and height of the image.

Structure from motion

images: The different pictures taken from a moving camera. There are ten pictures with name 0000.png to 0009.png.

keypoints: All positions of the keypoints for each image. Each row of a file is a tuple which gives the position of the keypoint on the image:

`--keypoints
   |--0000.png.txt
   |-- ...
   |--0009.png.txt

matches: The matches between keypoints for a pair of images. For each pair (000i.png, 000j.png), 000i.png-000j.png.txt contains the list of all keypoints and their corresponding index in the keypoints file of the images. For each keypoint, 000i.png-000j.png.txt gives a pair (idx_i, idx_j) where idx_i is the index of the keypoint in 000i.png.txt and idx_j is the index of the keypoint in 000j.png.txt.