We validate our method on four standard semantic segmentation benchmarks: CamVid, ADE-Bed, Cityscapes, and PASCAL Context. This directory contains the preprocessing scripts for each dataset. Please download the raw datasets using the links below and run the corresponding Python scripts.

Download:

• Dataset: SegNet-Tutorial repository

Usage:

python preprocess/camvid.py \
    --input_dir /path/to/SegNet-Tutorial/CamVid \
    --output_dir datasets/CamVid_256

--input_dir: Path to the downloaded original CamVid dataset
--output_dir: Path where preprocessed data will be saved (final files are saved to {output_dir}/real/train and {output_dir}/real/test)

We curated this dataset by selecting images from the bedroom category of the ADE20K dataset.

Download:

• Dataset: ADE_Bed_256.tar.gz

Usage:

tar -xzvf ADE_Bed_256.tar.gz -C datasets/
rm ADE_Bed_256.tar.gz

Download:

• Labels: gtFine_trainvaltest.zip (241MB)
• Images: leftImg8bit_trainvaltest.zip (11GB)

Extract the downloaded zip files before running the preprocessing script.

Usage:

The --input_dir should point to the Cityscapes root directory that contains gtFine and leftImg8bit folders:

Cityscapes/
├── gtFine/
└── leftImg8bit/

python preprocess/cityscapes.py \
    --input_dir /path/to/Cityscapes \
    --output_dir datasets/Cityscapes_256

--input_dir: Path to the extracted Cityscapes dataset directory (should contain leftImg8bit and gtFine folders)
--output_dir: Path where preprocessed data will be saved (final files are saved to {output_dir}/real/train and {output_dir}/real/val)

Download:

• Images: PASCAL VOC 2012 (Kaggle)
• Labels: trainval.tar.gz

Extract the downloaded files before running the preprocessing script.

Directory Structure:

After extracting pascal2012.zip:

pascal2012/
└── VOC2012/
    ├── Annotations/
    ├── ImageSets/
    ├── JPEGImages/
    ├── SegmentationClass/
    ├── SegmentationObject/
    └── VOC2012/

After extracting trainval.tar.gz:

trainval/
├── labels.txt
└── trainval/

The pascal_train.txt and pascal_val.txt split files are provided in the preprocess/ directory.

Usage:

python preprocess/pascal_context.py \
    --image_dir /path/to/VOC2012/JPEGImages \
    --mat_dir /path/to/trainval/trainval \
    --train_txt preprocess/pascal_train.txt \
    --val_txt preprocess/pascal_val.txt \
    --output_dir datasets/Pascal_context_256

--image_dir: Path to PASCAL VOC 2012 JPEG images directory
--mat_dir: Path to extracted trainval directory (should contain .mat files)
--train_txt: Path to train split file (provided in preprocess/pascal_train.txt)
--val_txt: Path to validation split file (provided in preprocess/pascal_val.txt)
--output_dir: Path where preprocessed data will be saved (final files are saved to {output_dir}/real/train and {output_dir}/real/val)

• All scripts resize images to 256x256 size.
• Labels are converted and saved in .npy format.
• Ignore labels for each dataset are converted to 255.