Domain-Based Semi-Supervised Learning: Exploiting Label Invariance in Unlabeled Data From Distributed Cameras

Leonardo Taccari; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 290-297

In several practical supervised learning problems where we have a large amount of data from distributed cameras or sensors, we can use domain knowledge to identify subsets of unlabeled examples with the same (unknown) label. Under this assumption, we propose a straightforward way to exploit label invariance in unlabeled data within a domain-aware semi-supervised learning framework (DSSL). Our approach exploits such invariance to generate higher quality pseudolabels to be used in a consistency loss term. We report experiments and ablation studies on three practical cases on data from real-world fleets of connected vehicles that naturally exhibit the required assumption: an image classification problem, a semantic segmentation task, and a time series classification one. We show that our approach is extremely effective, especially when few labeled samples are available, and can be easily adapted to tasks of different nature.