---
license: apache-2.0
datasets:
- yaraa11/enhanced-road-segmentation-dataset
language:
- en
pipeline_tag: image-segmentation
tags:
- adas
- autonomous-driving
- unet
- lane-detection
- semantic-segmentaion
- computer-vision
---


# 🛣️ Road & Lane Semantic Segmentation (U-Net with ResNet-50 Encoder)

## Model Description

This model performs **multi-class semantic segmentation for driving scenes**, focusing on:

- **Background**
- **Drivable road area**
- **Lane markings**

It is designed as a perception module similar to those used in **ADAS and autonomous driving pipelines**, where structured lane visualization is derived directly from segmentation outputs (no classical lane detection or Hough transform is used).

The model follows a **U-Net–style encoder–decoder architecture** with a **ResNet-50 backbone** pre-trained for feature extraction.

---

## Architecture

- Encoder: ResNet-50 (pre-trained)
- Decoder: U-Net–style upsampling path with skip connections
- Output: Softmax over 3 classes  
  - `0`: Background  
  - `1`: Road  
  - `2`: Lane markings  

---

## Intended Use

This model can be used for:

- Academic research in road scene understanding
- ADAS perception experiments
- Lane visualization systems based on segmentation
- Educational projects in computer vision and deep learning

### Limitations

- Not intended for real-world autonomous driving deployment
- Performance may degrade under:
  - Night conditions
  - Heavy rain or fog
  - Unusual camera angles
- Fine-tuning is recommended for different countries, road types, or camera setups

---

## Repository

Full training, fine-tuning, and inference pipelines are available [here](https://github.com/yaraeslamm/lane-detection)

---

## Training Data

The model was trained on a **custom enhanced dataset**, originally based on:

- **Semantic Segmentation Makassar (IDN) Road Dataset**  
  (~374 labeled images)

### Dataset Enhancements

To improve generalization, the dataset was expanded using:

- Random rotations
- Horizontal flipping
- Brightness and contrast jitter

Final dataset size: **~1,496 images**

The dataset is publicly available and not included in this repository due to size constraints.

---

## Training Procedure

- Loss Function: Sparse Categorical Crossentropy
- Optimizer: Adam
- Learning Rate: `1e-4`
- Epochs: 20
- Metrics:
  - Sparse Categorical Accuracy
  - Mean IoU
  - Lane-class IoU

Multiple experiments were conducted with different augmentation strategies and training schedules.  
Final model selection prioritized **lane IoU stability and visual consistency**, not only numerical metrics.

---

## Class Mapping

| Class ID | Label        |
|--------:|-------------|
| 0 | Background |
| 1 | Road |
| 2 | Lane Markings |

---

## Evaluation Results

| Metric | Training | Validation |
|--------|----------|------------|
| Accuracy | 0.9972 | 0.9954 |
| Mean IoU | 0.9456 | 0.9401 |
| Lane IoU | 0.8517 | 0.8542 |
| Loss | 0.0069 | 0.0146 |

---

## How to Use

This model is intended to be used with custom inference pipelines.

Typical inference steps:

1. Resize input image to model input size
2. Normalize pixel values
3. Run forward pass
4. Apply `argmax` over softmax output to get class IDs
5. Visualize lane pixels or overlay segmentation mask

### Example (TensorFlow / Keras)

```python
import tensorflow as tf
import cv2
import numpy as np

# -------- Load model --------
model = tf.keras.models.load_model(
    "model_path",
    compile=False
)

# -------- Load & preprocess image --------
img_path = "image_path"

orig = cv2.imread(img_path)
orig = cv2.cvtColor(orig, cv2.COLOR_BGR2RGB)

img = cv2.resize(orig, (256, 256))
img_norm = img / 255.0
img_input = np.expand_dims(img_norm, axis=0)

# -------- Predict --------
pred = model.predict(img_input)
mask = np.argmax(pred[0], axis=-1)  # (256, 256)

# -------- Create color mask --------
# Class mapping:
# 0 = background, 1 = road, 2 = lane

color_mask = np.zeros((256, 256, 3), dtype=np.uint8)

color_mask[mask == 1] = (255, 0, 0)   # Road -> Red
color_mask[mask == 2] = (0, 255, 0)   # Lane -> Green

# -------- Overlay full segmentation --------
overlay_full = cv2.addWeighted(img.astype(np.uint8), 0.6, color_mask, 0.4, 0)

# -------- Lane-only overlay --------
lane_mask = np.zeros_like(color_mask)
lane_mask[mask == 2] = (0, 255, 0)

overlay_lane = cv2.addWeighted(img.astype(np.uint8), 0.7, lane_mask, 0.3, 0)

# -------- Show results --------
cv2.imshow("Original", cv2.cvtColor(img.astype(np.uint8), cv2.COLOR_RGB2BGR))
cv2.imshow("Segmentation Overlay (Road + Lane)", cv2.cvtColor(overlay_full, cv2.COLOR_RGB2BGR))
cv2.imshow("Lane Only Overlay", cv2.cvtColor(overlay_lane, cv2.COLOR_RGB2BGR))

cv2.waitKey(0)
cv2.destroyAllWindows()


```

For full pipelines including video processing and lane visualization,  
see the [GitHub repository](https://github.com/yaraeslamm/lane-detection)


---

## Fine-Tuning

Fine-tuning is recommended if:

- Using different road environments
- Working with different camera perspectives
- Wanting to rebalance lane vs road classes

Training scripts support:

- Freezing the encoder
- Full model retraining
- Custom datasets with RGB masks converted to class IDs

---

## Ethical Considerations

This model is intended for **research and educational purposes only**.

It should not be used as a sole perception system in safety-critical or real-world autonomous driving applications.

---

## License

This model is released under the **Apache 2.0 License**, allowing commercial and research use with attribution.

---

## Author

Developed by Yara Elshehawi  
Check out my other work:
[Portfolio](https://yaraeslamm.github.io)
[LinkedIn](https://www.linkedin.com/in/yara-eslam-877421212/)