Semantic Segmentation Competition

This repository contains code and results for a semantic segmentation competition using a small autonomous driving dataset. The project explores various data augmentation strategies, network architectures, and training techniques to improve segmentation accuracy and efficiency on urban road scenes.

Dataset

• Training images: 150
• Testing images: 50
• Classes: 19 (see seg_data/names.txt)
• Color palette: see seg_data/colors.txt

Installation

pip install -r requirements.txt

Project Structure

.
├── segmentation-2025.ipynb          # Main notebook (all experiments consolidated)
├── requirements.txt                  # Python dependencies
├── Final_Report.pdf                  # Detailed report with ablation studies
├── seg_data/
│   ├── training/image/               # Training images
│   ├── training/label/               # Training labels
│   ├── testing/                      # Testing images
│   ├── names.txt                     # Class names
│   └── colors.txt                    # Class color palette
└── result/
    ├── baseline/                     # Baseline model
    ├── flip_augmentation/            # Flip augmentation experiment
    ├── flip_colorjitter_batchnorm/   # Flip + color jitter + batch normalization
    ├── baseline_augmentation/        # Baseline with data augmentation
    ├── residual_skip_dilated_augmentation/  # Residual + skip + dilated convolutions
    ├── mobilenetv3_crossentropy/     # MobileNetV3 with cross-entropy loss
    ├── mobilenetv3_hybrid_loss/      # MobileNetV3 with hybrid (CE + Dice) loss
    ├── mobilenetv3_augmentation_crossentropy/  # MobileNetV3 + augmentation + CE loss
    ├── crop_rot_blur_jitter_hybrid_loss/  # Crop + rotation + blur + jitter + hybrid loss
    └── crop_rot_blur_jitter_adamw_batchnorm_residual_skip_dilated/  # Full pipeline experiment

Usage

1. Prepare Data

   • Place the dataset in the seg_data/ directory following the structure above.

2. Run Experiments

   • Open and run segmentation-2025.ipynb for the main experiment pipeline.
   • Each subfolder in result/ contains a notebook for a specific ablation study.

3. Evaluate

   • Evaluation uses mean Intersection over Union (mIoU) as the primary metric.
   • Results and predictions are saved in the result/ subfolders.

Results and evaluation scope

The main notebook’s historical ablation table lists baseline mIoU 0.286 / 66.97 GFLOPs, a residual/skip/dilated configuration at 0.3421 / 204.83 GFLOPs, and a MobileNetV3 configuration at 0.3377 / 18.10 GFLOPs. Values are recorded notebook results, not a new rerun or an estimate across seeds. The table header says percent, but its listed mIoU values use the 0–1 scale.

The checked-in training loop measures testing/label during training and uses that score for best-checkpoint selection and early stopping. Consequently, the directory called testing is a development/selection set in this workflow; these results are not an untouched final-test claim. A future study needs a separate validation partition and a frozen final test before making generalization claims.

See the main notebook, the report and individual experiment notebooks for historical settings. Alternative architecture cells are present; choose the matching model, loss and training/prediction path rather than assuming every cell forms one universally runnable pipeline.

References

• PyTorch Documentation
• fvcore FLOPs Analysis