This repository contains code written for a competition in vanishing point detection in images using Hough transform.

The task of detecting the vanishing point in images is of great importance and has a wide range of applications. It arises in the development of systems for autonomous vehicles (localization on the road surface, self-calibration), in the analysis of scene lighting (analysis of color histograms), and in correcting projective distortions.

800 labeled images of the road surface from a vehicle camera, containing coordinates of the vanishing point, are used as data for constructing and testing the algorithm. During testing, the images are augmented by rotation.

As a result, the algorithm returns a JSON file in the format:

{
    "file1.jpg": [x1, y1], 
    "file2.jpg": [x2, y2]
}

The test sample is generated using the script test_generation/test_generation.py:

python test_generation/test_generation.py --s path_to_dataset --d path_to_save_new_dataset --num num_of_imgs_to_generate --seed seed

It is assumed that the data folder is structured as follows:

├── dataset
    ├── markup.json
    ├── source

The quality of the solution is assessed using an angular metric. The proximity of the predicted point $A$ to the true value $B$ is determined by the angle $\alpha$ between the vectors $\xi$ and $\eta$ drawn to these points from the point $O$, as demonstrated in the figure below.

$$\alpha = \arccos \frac{\left \langle \xi, , \eta \right \rangle}{|\xi||\eta|},$$

where