Online Domain Adaptation for Semantic Segmentation
in Ever-Changing Conditions
ECCV 2022

  • Theodoros Panagiotakopoulos
  • Pier Luigi Dovesi
  • Linus Härenstam-Nielsen
  • Matteo Poggi
OnDA (literally "wave" in italian) allows for adapting across a flow of domains, while avoiding catastrophic forgetting.

Abstract

Teaser

Unsupervised Domain Adaptation (UDA) aims at reducing the domain gap between training and testing data and is, in most cases, carried out in offline manner. However, domain changes may occur continuously and unpredictably during deployment (e.g. sudden weather changes). In such conditions, deep neural networks witness dramatic drops in accuracy and offline adaptation may not be enough to contrast it. In this paper, we tackle Online Domain Adaptation (OnDA) for semantic segmentation. We design a pipeline that is robust to continuous domain shifts, either gradual or sudden, and we evaluate it in the case of rainy and foggy scenarios. Our experiments show that our framework can effectively adapt to new domains during deployment, while not being affected by catastrophic forgetting of the previous domains.

Method overview

Method

Our framework OnDA is designed to deal with Online Domain Adaptation across multiple domains. While adopting state-of-the-art UDA strategies for prototypical self-training, we design a teachers-student approach which allows for dynamic orchestration of the teachers. The adaptation process is guided by the nature of the domain change itself. Indeed, our method actively chooses the best teacher to employ for training the student model, according to the domain shift intensity and direction. The core of our teachers switching policies consists of the continuous monitoring of the domain shifts happening during deployment.

Citation 

    @inproceedings{Panagiotakopoulos_ECCV_2022,
    title     = {Online Domain Adaptation for Semantic Segmentation in Ever-Changing Conditions},
    author    = {Panagiotakopoulos, Theodoros and
                 Dovesi, Pier Luigi and
                 H{\"a}renstam-Nielsen, Linus and
                 Poggi, Matteo},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year = {2022}
    }

Acknowledgements

The authors thank Hossein Azizpour, Hedvig Kjellström and Raoul de Charette for the helpful discussions and guidance.