This project focuses on semantic segmentation of road scene images using the BDD100K dataset.
A baseline segmentation model is developed using a MobileNetV2-based architecture trained on labeled data.
To overcome the limitation of labeled data, a pseudo-labeling approach is applied, where the trained model generates labels for additional data. High-confidence predictions are selected and combined with the original dataset to retrain and improve the model.
The performance is evaluated using pixel accuracy and mean Intersection over Union (mIoU), showing improvements in segmentation quality.