pretrained weights prediction have a little bias? about lstr

[liuruijin17]

Hi,
I think the reasons may consist of (1) lack of appearance clues at the remote horizon; (2) lack of aforementioned data.

[lucasjinreal]

@liuruijin17 Did u notice these samples which happens mostly in curved scenarios? Is there a way to make LSTR solve it? (Otherwise this problem was quite same as UltraFfastLane algorithm which based on anchor, the lane line can not locate on lane mark tightly)

[liuruijin17]

Hi @jinfagang
We have noticed these samples and here are some immature ideas:

• The number of labeling points at the remote horizon is very sparse.
• The data imbalance between points at the curved regions and linear aligned points.
• The curvature distribution of point is hard to balance, even though the number of curved lanes can be ensured to collect a sufficient number of training lanes.
• The appearance information at the remote horizon is severely compressed due to projective transformation.

How to solve it?

• New training loss which can balance the data imbalance.
• Novel architecture to focus on the remote region.
• More suitable mathematical modeling for the lane structures at the remote region.

[lucasjinreal]

@liuruijin17 Waiting for LSTRv2.0

[liuruijin17]

@jinfagang Thank you for your interest and support, and it is a pleasure to discuss this with you!