[D] How can I detect the text orientation using MMOCR or MMDET models?

My training images have texts that appear in various orientations on the image. As a result, I don't know what's their original orientation since for example DBNetPP does not return the bbox angles in the corners in a natural orientation order. How can I solve this issue? I have tried other pretrained detection models, but they also does not do that, maybe because they were not trained on rotated images. How can I solve this issue?

https://preview.redd.it/tvq6fp9k3zxc1.png?width=1000&format=png&auto=webp&s=ecf3f3e757e6450e34c1257f9eb8e0fec4ce7bba

https://preview.redd.it/yea66hdl3zxc1.png?width=1000&format=png&auto=webp&s=4eafb6d4354c6a0d851d6b5fad456f99441d9bc2

example image images issue machinelearning natural solve text training