Lane Detection Models: BézierLaneNet, MobileNets, RepVGGs & Swin Transformer

PytorchAutoDrive v3.1 (2022Q1 Release)

IMPORTANT: This is not a Happy April Fools Day.
This release is mostly about adding new models for lane detection, and now we have a paper reference! Although not all datasets & method variations are tested for the new backbones, you can do that fairly easily with the current config-based coding style.
There are in total 12 closed issues, 9 merged PRs and 14 commits to master branch since v3.0.

Contributors: @voldemortX @cedricgsh
Thanks @FengqiLiu1221 @junshutang for generously providing GPUs!

Major features

1. A series of RepVGG backbones are added for lane detection Baseline and SCNN (CULane). #54
2. Swin-Tiny is supported as backbone for lane detection Baseline (CULane). #56
3. MobileNetV2 and MobileNetV3-Large are supported as backbone for lane detection Baseline and RESA (CULane, TuSimple). #53
4. BézierLaneNet (ResNet-18, ResNet-34) is supported for lane detection (CULane, TuSimple, LLAMAS). #60
   With this merge of the private repo, we now have:
   • A mature system for keypoint loading & transforms (including a new data augmentation policy level 1b), see functional_keypoints.py for useful functions & implementation details.
   • The Cosine Annealing learning rate schedule.
   • Polynomial & Bézier curves supports (lane fitting, upper-bound test, etc.), see CURVE.md for more info.
   • A way to implement customized flops count for customized layers. An example.

Minor features, Docs, Refactors & Bug fixes

1. Per-model documentation as suggested in #62 are started, part of them are finished, others pending (we would love the community to help with this). #63 #64 #67 #69
2. The Arxiv release of the paper. #61

Known BC-Breaks

1. A silent BC-Break (more of an improvement). The LSTR CULane models are re-evaluated with threshold 0.95 to align with BézierLaneNet, the performance is slightly increased (~ 1%), while the downloaded weights remain the same. #60
2. The old simple and strong augmentation policies for lane detection are renamed to level 0 and level 1a to accommodate more new augmentation policies. It is simply a name change, no trained models will be affected. #60