Lane line detection combining a bird's-eye-view edge proposal network with Swin Transformer global context modelling on TuSimple highway images.
Two variants — sequential and parallel — were implemented and ablated to resolve an architectural ambiguity in the original paper. The sequential variant is recommended: statistically equivalent accuracy, 25% faster.
| Model | Accuracy | TPR | FPR | F1 | FPS |
|---|---|---|---|---|---|
STLaneNet_Seq |
91.86% | 80.17% | 12.18% | 83.82% | 162.9 |
STLaneNet_Par |
91.75% | 80.20% | 12.64% | 83.63% | 130.2 |
| Paper (Du et al., 2024) | 98.85% | 97.5% | 3.8% | ~0.94 | 64.8 |
Reproduction verdict: Architecture faithfully reproduced. The 6.99 pp accuracy gap is a training-step deficit (4,050 vs 80,000 steps), not an architectural failure. FPS claim surpassed 2.5× (162.9 vs 64.8). The sequential/parallel ambiguity in the paper resolves empirically in favour of STLaneNet_Seq.
ST-LaneNet is a dual-branch network operating at 368×640 resolution.
Branch 1 — Edge Proposal (bird's-eye view)
Image → IPM warp (H) → SpatialPriorHead → binary mask
Image/mask → EdgeEncoder (depthwise-separable, dilated) → CARAFE decoder → warp back (H⁻¹)
Branch 2 — Localization (front-view)
Image → Swin-Tiny (4 stages) → loc_reduce (768→128) → bilinear upsample
Fusion head: concat [edge 64ch, loc 128ch] → 192ch → Conv → 1×368×640 lane logit.
Total: ~29,656K parameters (Swin-Tiny alone = 95.4%).
→ Full architecture details, layer-by-layer table, parameter count
pip install marimo openmimDataset is auto-downloaded from Kaggle on first run. Set Kaggle credentials:
# Option 1: file
echo "KGAT_..." > ~/.kaggle/access_token
# Option 2: env vars
export KAGGLE_API_TOKEN=KGAT_...
export KAGGLE_USERNAME=your_usernameTraining + Evaluation (combined notebook):
marimo run notebook.py # headless
marimo edit notebook.py # interactive / debugThe notebook trains both variants sequentially (Par then Seq), then runs evaluation with GT-guided metrics and saves figures to report_figures/.
Load pre-trained weights instead of training:
Weights are pulled automatically from Kaggle at the start of the evaluation section if ./ST_LaneNet_Weights/STLaneNet_Par_best.pth and STLaneNet_Seq_best.pth are not present.
notebook.py # Marimo notebook: training + evaluation (canonical source)
notebook.ipynb # Jupyter export from Marimo platform (read-only snapshot)
report_figures/ # Auto-generated by notebook.py (figures for the report)
stln-par/{1,5,10,...}.png # Par model training progression snapshots
stln-seq/{1,5,10,...}.png # Seq model training progression snapshots
ST_LaneNet_Weights/ # Checkpoint directory (created on first training run)
docs/ # Extended documentation (this repo)
| File | Contents |
|---|---|
| docs/related-work.md | Traditional methods → deep learning; why ST-LaneNet |
| docs/architecture.md | IPM, SpatialPriorHead, DSConv math, CARAFE, Swin Transformer, Par vs Seq, layer-by-layer table, parameter count |
| docs/critical-analysis.md | Flaws in the original paper: FC-1000, fusion mislabelling, IPM absence, Seq/Par contradiction, Swin confusion, 80k epochs typo |
| docs/training.md | Focal Loss derivation, hyperparameters, GT mask generation, bfloat16 vs float16, .detach() constraint, gradient clipping, infrastructure |
| docs/results.md | GT-guided evaluation algorithm, ablation table, post-processing, SOTA comparison, training curves, qualitative analysis, limitations, ethics |
| Resource | Link |
|---|---|
| TuSimple dataset (Kaggle) | https://www.kaggle.com/datasets/manideep1108/tusimple |
| Trained weights (Kaggle) | https://www.kaggle.com/datasets/youtikig/st-lanenet-weights-backup |
Y. Du, R. Zhang, P. Shi, L. Zhao, B. Zhang and Y. Liu,
"ST-LaneNet: Lane line detection method based on Swin Transformer and LaneNet,"
Chinese Journal of Mechanical Engineering, vol. 37, 2024.
https://doi.org/10.1186/s10033-024-00992-z