Model Card for lane-detection-unet-ncnn

This model card documents the UNet-based lane segmentation models and NCNN deployment pipeline for the BDD100K dataset.

Model Details

Model Description

Developed by: Nick Pai
Model type: UNet, UNetDepthwise, UNetDepthwiseSmall, UNetDepthwiseNano (semantic segmentation)
Language(s): Not applicable (computer vision)
License: MIT
Finetuned from model: Custom UNet architectures

Model Sources

Repository: https://github.com/nick8592/lane-detection-unet-ncnn
Paper: None (see repo for references)
Demo: See README for C++ and Python inference examples

Uses

Direct Use

Lane segmentation for autonomous driving, road scene analysis, and research
Python and C++ inference (PyTorch, NCNN)

Downstream Use

Can be fine-tuned for other road segmentation tasks or datasets

Out-of-Scope Use

Not suitable for non-road or non-lane segmentation tasks
Not designed for medical or non-automotive domains

Bias, Risks, and Limitations

Trained on BDD100K, so performance may degrade on out-of-distribution scenes
Lane annotations may contain labeling errors or bias
Not robust to extreme weather, occlusions, or unusual camera angles

Recommendations

Users should validate model performance on their own data before deployment
Consider retraining or fine-tuning for new domains

How to Get Started with the Model

See the README for setup, training, and inference instructions. Example:

# Python inference
python3 scripts/test.py

Training Details

Training Data

BDD100K lane segmentation subset
Images: 100k train/val/test split
Masks: binary lane masks

Training Procedure

Standard UNet training with Adam optimizer
Image size: 256x256
Batch size: 8-32
Epochs: 10-20
Augmentation: random flip

Training Hyperparameters

fp32 precision
Learning rate: 1e-3 (default)

Speeds, Sizes, Times

Training time: ~10 hours on RTX 4060 for 10 epochs

Evaluation

Testing Data, Factors & Metrics

BDD100K val set
Metrics: IoU, Dice/F1, Pixel Accuracy, MAE

Model Examination

Visualizations and overlays available in README and assets/

Technical Specifications

Model Architecture and Objective

UNet encoder-decoder, depthwise variants for efficiency
Objective: binary lane segmentation

Compute Infrastructure

Hardware: RTX 3090, Intel
Software: PyTorch 1.13, NCNN, OpenCV

Citation

BibTeX:

@software{Pai_UNet_Lane_Segmentation_2025,
author = {Pai, Nick},
license = {MIT},
month = sep,
title = {{UNet Lane Segmentation \& NCNN Deployment}},
url = {https://github.com/nick8592/lane-detection-unet-ncnn},
version = {1.0.0},
year = {2025}
}