One PyTorch toolbox for lanes and road pixels

What it does PytorchAutoDrive is a pure Python framework built on PyTorch that collects semantic segmentation models (FCN, DeepLab, ENet, ERFNet) and lane detection models (SCNN, RESA, LSTR, LaneATT, BézierLaneNet) under one roof. It handles training, testing, visualization, and benchmarking through config files rather than scattered scripts. The project also packages ONNX and TensorRT export so the same codebase can move from experiment to deployment without a rewrite.

The interesting bit The maintainers treat this as a fair benchmark as much as a model zoo: they claim their implementations train faster on a single GPU and often outperform other public versions, with technical notes in the wiki explaining why. That focus on reproducibility and modulated, readable code is rarer than it should be in vision research repositories.

Key highlights

• Config-based training and testing for all supported models, backbones, and datasets.
• Mixed precision training, TensorBoard logging, and built-in visualization for images and video.
• Benchmarking tools that measure FPS, FLOPs, and memory per model.
• Deployment support via ONNX and TensorRT export.
• Single-GPU training for all models; trained weights and shell scripts available in the Model Zoo.

Caveats

• A few items are explicitly unfinished: BDD100K support is marked in progress, SAD is marked postponed, and PRNet is also marked in progress.
• The README warns the repo is under active development, though uploaded model results are described as stable.

Verdict Researchers and engineers working on autonomous driving perception who want a unified training and benchmarking environment should look here. If you only need a single off-the-shelf model and do not care about comparing methods, a narrower repository might be less to digest.