2024-10-14 Sporta Centrs Task

TODO

1. Port from Python https://github.com/asya-ai/sporta_centrs_scoreboard_ocr (invite sent)

2. To C++: https://github.com/asya-ai/sporta_centrs_scoreboard_cpp (invite sent)

Create Visual Studio C++ project (latest version), if necessary can add dynamically linked libraries, but must have to use some exactly these versions

x
1
Qt 5.14.1
2
opencv 4.5.3
3
boost 1.83
4
openssl 1.1.1.d
5
cryptopp_700
6
bmd 11.5.1
7
ndi 6.0
8
aja 15.2.0.4
9
ffmpeg 4.4
10
libjpeg-turbo

Inplementation of function

Must implement Dynamic DLL with a function get_transformation_from_photo(cv::Mat rgb_photo) -> struct_output

struct_output must have cv::Mat transformation (opencv transformation matrix) and if possible to find then also List<cv::Vec4f > contour before transformation in input pixel space, contour can be split in multiple lines, and cv::Mat rgb_photo_transformed for verification and cv::Mat rgb_photo_original_with_contour original image with lines drawn on top of it for validation

Write test code in console app, where path to file given as argument and output is new transformed image.

There are 2 stages within get_transformation_from_photo:

1. Algoirthms are used to extract all lines in the photo

2. Algoirthms are used filter lines, improve contour and generate transformation matrix. If full or partial countour is found then also these lines can be used, otherwise looking for alternative ways to get transformation matrix

Assumptions:

1. Tablo is in the middle of photo

2. Tablo takes significant portion of photo

AI tools

To speed up implementation

1. Try adding in comments pyhon code and generate CPP code

2. Try commanding selected code to be ported using http://cursor.com (if necessary we will buy pro version)

3. perplexity.ai give example in C++ LibraryX how to do Z

Porting strategy

Document as much as possible code with references in comments to original python files

The main logic is based on:

1.) openCV → directly available in C++

2.) Numpy → Armadillo, Eigen (choose simplest ad most documented)

3.) Shapely → directly built on GEOS (original C++ library)

AI models:

1.) Segmentation model:

pytorch → TorchScript, ONNX format

2.) deepLSD:

https://github.com/cvg/DeepLSD

torch, numpy, scipy - only implement functions used, not all

Problematic part here might be the merging, filtering and scoring functions used to process the predicted lines from the model.

Full requirements can be found: https://github.com/cvg/DeepLSD/blob/main/requirements.txt

Trained model weights

Model weights must be in ONNX format, get latest from reinis

Lines of code:

Ours: 37 executed functions within 645 lines of code

deepLSD: 43 executed functions within 1061 lines of code

Coverage report:

File overview: