This project performs semantic segmentation on the CamVid dataset — a benchmark for autonomous driving scene understanding. The dataset used is slightly different from the original CamVid, with a different train-test split.
Implemented and compared some traditional segmentation models:
- Classical U-Net
- U-Net with ResNet backbones (ResNet-34, ResNet-50, ResNet-101)
- DeepLabV3+ with ASPP and dilated convolutions
Goals:
- Evaluate how architecture and backbone depth affect segmentation quality.
- Evaluate how useful pretrained backbones are.
- Effect of Focal loss.
| Property | Details |
|---|---|
| Domain | Road scenes / Autonomous driving |
| Labels | Pixel-wise semantic segmentation classes |
| Data split | 500/100/100 |
| Model | Backbone | Notes |
|---|---|---|
| U-Net | ResNet-50 | ⭐ Best overall performance |
| U-Net | ResNet-34 / ResNet-101 | Higher capacity → improved results |
| DeepLabV3+ | ResNet backbone | Improves with fine-tuning |
- Increasing model capacity improved segmentation accuracy
- Pretrained backbone implementation significantly outperformed custom training
- Class imbalance remains a major challenge. Adding weighted Focal loss doesn't improve results.
For a more comprehensive overview, please see the accompanying PowerPoint presentation.