Skip to main content

Reimplementing SCNN for Lane Detection: Architectural Improvements and TensorRT Deployment

· 3 min read
Yi-Chen Zhang
Yi-Chen Zhang
Lead Engineer, AI and Autonomous

I recently worked on a lane marking detection project. After reviewing the literature, I decided to use the SCNN (Spatial CNN) model — not because it's state of the art, but because it falls squarely into my "if you can't implement it, you don't really understand it" category.

SCNN is simple, manageable with limited GPU resources, and well-suited for understanding the fundamentals of lane detection. The video above demonstrates TensorRT GPU inference on the KITTI dataset.

Architectural Changes​

Compared to the original implementation, I made three significant changes:

1. Framework and Modularization​

The original implementation is written in Lua, which is difficult to follow without prior experience. I reimplemented the model in PyTorch and modularized all components, making the codebase cleaner and easier to reason about.

2. Segmentation (Decode) Head Redesign​

The original segmentation head uses simple interpolation for upsampling. I replaced it with an FCN-style decoder using transposed convolutions, which provides:

  • Smoother outputs
  • Better numerical stability
  • Improved gradient flow during training

3. Existence Head Redesign​

This is the most impactful change. The original existence head uses two fixed-size fully connected layers, which are unfriendly for ONNX and TensorRT export. I replaced it with a fully convolutional, FCOS-style classification head.

Lane existence is analogous to object vs. background classification, making the FCOS-style head a natural fit. This change:

  • Removes the fixed-size input constraint
  • Reduces parameter count significantly
  • Stabilizes existence probability predictions
  • Enables ONNX and TensorRT export for arbitrary image sizes

Why This Matters for KITTI​

CULane images are 590×1640 (resized to 288×800, aspect ratio ≈ 2.78), while KITTI images are 370×1226 (aspect ratio ≈ 3.31). The original fixed fully connected layers implicitly assume a fixed aspect ratio, causing geometric distortion when applied to KITTI images. The fully convolutional design avoids this issue entirely.

Results​

Despite being trained only on CULane — with no KITTI-specific training or fine-tuning — the model demonstrates strong generalization:

MetricImprovement
F1 score~1% improvement
Existence head parameters~1000× fewer
ONNX / TensorRT export✅ Resolution-agnostic
Validation categoriesBetter in 7 out of 8

Failure Cases​

Most failures occur in challenging scenarios that are largely data-driven:

  • Tunnel shadows — heavily shadowed lane markings
  • Missing or degraded lane lines — worn or absent road markings
  • Intersections — cross-traffic scenarios
  • Sharp curves — tight geometry at the edge of the training distribution
  • Strong sunlight — overexposure washing out lane markings

More aggressive or targeted data augmentation would likely mitigate these. Given the simplicity of the model and the complete absence of KITTI-specific training, I'm already very satisfied with the results.