search

cfzd / Ultra-Fast-Lane-Detection

Ultra Fast Structure-aware Deep Lane Detection (ECCV 2020)
MIT License
1.76k stars 491 forks source link

About row anchors #333

Open Tu-Ear opened 1 year ago

Tu-Ear commented 1 year ago

Hi @cfzd , thanks for your great work! I have some questions about row anchors： 1、Why the step of row anchors for two datasets are different and will the size of step influence the accuracy? 2、How did you decide the range of row anchors for different datasets? 3、I read the issue: #8 and I wonder if I increase the number of row anchors on CULane, can I set the sim_loss_w to 1 and improve the model? Thank you!

cfzd commented 1 year ago

@Tu-Ear

  1. This question can be found in our new version UFLDv2: https://arxiv.org/pdf/2206.07389. Generally speaking, the more anchors, the higher performance.
  2. The range should cover the possibale range of lanes.
  3. You can give it a try. In UFLDv2, we designed a new expectation loss to replace structural loss.