Open voldemortX opened 2 years ago
Hi, thanks for your interest and contributions!
The CULane is a bit more difficult since their lane points at occluded regions are somewhat too casual. In some scenes, there are no visible lanes, but there are still 'imaged' lanes, making their annotations not consistent with the image appearances. Maybe that's the reason all lane detectors perform well on CULane have relatively larger backbones to hold and 'memorize' clearly.
However, LSTR still has lots of potentials to be improved. I had finished my other projects recently, thus I would like to bring some updates to LSTR.
Again, thanks for your contributions!
@liuruijin17 Hi! I've trained LSTR on CULane. A ResNet18 reduced version as in this repo achieved 67 F-score, while a standard ResNet34 achieved 71.52 at a lower FPS (~60). The epochs needed for convergence is 150.
It would seem that LSTR performance looks decent on CULane as well as TuSimple.
Details are here.