Foggy Road Scene Segmentation

Fine-tuning SegFormer-b0 for semantic segmentation under foggy conditions using the ACDC adverse weather dataset.

Results

Model	mIoU
Baseline (Cityscapes only)	54.5%
Fine-tuned (+ ACDC Fog)	60.6%

+11.1% improvement after fine-tuning on 400 foggy driving images.

Qualitative Comparison

mIoU Results

What this project does

• Loads a pre-trained SegFormer-b0 model trained on Cityscapes
• Evaluates its baseline performance on foggy road scenes from ACDC
• Fine-tunes it on 400 foggy training images for 10 epochs
• Evaluates and compares mIoU before and after fine-tuning
• Visualizes segmentation predictions using the Cityscapes color map

Dataset

ACDC — Adverse Conditions Dataset with Correspondences Fog split: 400 train / 100 val images with pixel-level semantic labels.

Model

SegFormer-b0 fine-tuned on Cityscapes by NVIDIA, used as the baseline and starting point for fine-tuning.

Tech Stack

• Python, PyTorch
• HuggingFace Transformers (SegFormer)
• OpenCV, Matplotlib, NumPy
• Google Colab (T4 GPU)

How to run

1. Open notebooks/foggy_segmentation.ipynb in Google Colab
2. Mount your Google Drive and point paths to your ACDC dataset
3. Run all cells top to bottom