Key Points Estimation and Point Instance Segmentation Approach for Lane Detection

RONELDv2: A faster, improved lane tracking method

“…More recently, Qin et al [21] introduced a method which treats the lane detection process as a row-based selection problem using global features, achieving compelling results with the ResNet-34 model on the CULane [8] dataset while running around 23 times faster than SCNN. Other methods have also approached the lane detection problem as an instance segmentation task [30], [31] or model lanes in three dimensions [23], the latter which has achieved competitive results even on the TuSimple dataset [7], an image-only lane detection dataset. Some Generative Adversarial Networks (GANs) [32] have also been recently introduced to address the lane detection problem [33], [34], where the network benefits from seeing both real and generated fake predictions at the same time to improve lane detection results.…”

A Survey of Localization Methods for Autonomous Vehicles in Highway Scenarios

“…In Table 4 and Table 5 , we provide an overview of the five best performing methods with available code on each benchmark, at the time of writing of this manuscript. Compared to the available tables on , we added whether extra training data were required in the CULane benchmark, and added the reference [ 103 ]. The difference between the top algorithms in both tables is only by a few percentages, whether it is based on the accuracy score in the Tusimple benchmark, or based on the F1 score in both benchmark.…”

Model-Agnostic Defense for Lane Detection against Adversarial Attack

“…We achieve accuracy within 2% of Neven et al's results before adding our defense. Note that due to the lane detection model-agnostic nature of our defense, the results from our experiments should be applicable to any other model we could have chosen, such as [5], [6], and [7].…”