Code: MemCDT

Domain-Adaptive Semantic Segmentation with Memory-Efficient Cross-Domain Transformers

This code release contains the implementation of the Unsupervised Domain Adaptation (UDA) framework for Transformer-based Semantic Segmentation models described in the paper listed below. Building upon state-of-the art self-training UDA schemes, the framework exploits a novel three-branch Transformer architecture that combines self- and cross-domain attention mechanisms for better source-target feature alignment. The key aspect of the developed approach lies in its ability to yield enhanced adaptation capabilities compared to its baselines, without increasing the GPU memory footprint during training.

The software for the UDA training pipeline is publicly available and can be accessed here.

Users of this software are kindly asked to cite this paper, where it was introduced:

 

Ruben Mascaro, Lucas Teixeira, and Margarita Chli, "Domain-Adaptive Semantic Segmentation with Memory-Efficient Cross-Domain Transformers", in Proceedings of the British Machine Vision Conference (BMVC), 2023. Research Collection Presentation Video