about the inference results?(have picture)

[Onehundred0906]

i use the model which is trained by tusimple datasets, and when i inference the picture that i download from website with detect.py, the detect results is not correct, and i have resize the pic as 720x1280 before inference and set the conf_threshold as 0.01.please help explain and can give me some suggestions to improve it.

here is my personal thinking, because the anchor is learned, so the downloaded pic's anchor is not suitable for tusimple pic, so if i want a good result, i must use the downloaded pic to train?