How does the model perform on CULane datasets? about lstr HOT 4 CLOSED

In next week. @lunachy

from lstr.

Thanks!
I have trained a model on the CULane dataset recently, with the same configuration (no additional parameters) the model achieves 65.6 F1-measure. After multiplying all channels by 2, the F1-measure reaches 68.6.

However, I think my evaluation code might have some problems because the original CULane evaluation protocol considers the segmentation to be the input. I directly generate predicted sequences using my own python codes by depending on their Matlab codes as I understand it (which might have some problems).

I will update my codes about CULane in the future, but anyway please feel free to pull requests!

from lstr.

Thanks!
I have trained a model on the CULane dataset recently, with the same configuration (no additional parameters) the model achieves 65.6 F1-measure. After multiplying all channels by 2, the F1-measure reaches 68.6.

However, I think my evaluation code might have some problems because the original CULane evaluation protocol considers the segmentation to be the input. I directly generate predicted sequences using my own python codes by depending on their Matlab codes as I understand it (which might have some problems).

I will update my codes about CULane in the future, but anyway please feel free to pull requests!

any plan about pushing CULane dataset‘s work?

from lstr.

I have uploaded training testing and evaluating codes about culane in the culane branch, please check it!
Now, I will close this issue.

from lstr.