Constant wrong prediction using LETR + line scoring

[valentin-fngr]

Hi,

After installing all packages and setting the project up, I finally run the following command :

PYTHONPATH=$PYTHONPATH:./LETR/src python modelFitting_letr.py --checkpoint-filepath [last stage checkpoint filepath] --img [image filepath] --output_path [dirpath where save the result]

And here are my detection result :

NOTE : I did not retrain anything, I did not perform transfer learning etc.

Am I doing something wrong or is the model pure total garbage ?

[valentin-fngr]

One more :

[TEXflip]

Hi, the model it works in a small validation set and it covers a small domain space (this project have been made for a small university project, it's not meant for commercial quality level). In this case it is not working because it can't see enough court lines, but the algorithm it can be easily improved.