Project crafted by Antonio Ferrigno, Giulia Di Fede and Vittorio Di Giorgio for the Advanced Machine Learning course at Politecnico di Torino (2023/2024)
In this work, we tackle the challenging task of real-time domain adaptation in semantic segmentation. We experiment on a novel and efficient architecture, the Short-Term Dense Concatenate (STDC) network, for semantic segmentation. We combine this with adversarial learning to align the feature distributions of the source and target domains, as specified in Learning to Adapt Structured Output Space for Semantic Segmentation.
We also explore several extensions to improve the performance and efficiency of the domain adaptation process. These include:
- Depthwise Separable Convolution
- Fourier Domain Adaptation
- Self-supervised Training with Multi-band Transfer.
We conduct extensive experiments on two datasets, GTA V and Cityscapes. Our results demonstrate significant advancements in reducing the domain gap and enhancing the segmentation accuracy in real-time scenarios.
