Abstract
Unsupervised Domain Adaptation (UDA) for semantic segmentation often suffers from "confirmation bias" during self-training, where models reinforce their own incorrect predictions. This paper presents EPS-UDA, a framework that utilizes a tri-learning architecture to provide implicit pseudo-supervision. By employing three parallel classifiers that generate consensus-based pseudo-labels for each other, the system breaks the error-reinforcement loop. Combined with Semantic Feature Alignment (SFA) and Adaptation Ability Estimation (AAE), EPS-UDA achieves state-of-the-art results on major synthetic-to-real benchmarks like GTA5-to-Cityscapes.
Introduction
Training deep learning models for semantic segmentation requires massive amounts of pixel-level manual labeling, which is prohibitively expensive. A common solution is Unsupervised Domain Adaptation (UDA), where models are trained on synthetic data (e.g., video games) and adapted to real-world images.
However, standard self-training approaches often fall into a confirmation bias trap: the model makes an error, generates a confident but incorrect pseudo-label, and then trains itself to repeat that mistake. Additionally, the "domain gap" between synthetic and real textures frequently causes feature collapse, where different semantic classes become indistinguishable in the feature space.
Methodology
The authors propose EPS-UDA, which rests on three primary technical pillars designed to enhance the reliability of pseudo-labels and the quality of feature alignment.
1. Tri-learning Architecture
Instead of a single model learning from its own predictions, EPS-UDA uses three segmentation networks ($F_1, F_2, F_3$) sharing a common backbone. In a "two-against-one" setup, the consensus of two branches provides pseudo-labels to train the third. This diversity ensures that no single network's errors dominate the training signal, providing complementary knowledge rather than reinforcing existing biases.
2. Semantic Feature Alignment (SFA)
EPS-UDA treats foreground and background categories differently to account for their distinct geometric properties:
- Background (Stuff): Large, consistent regions like roads or sky are aligned by minimizing the distance between feature centroids of the source and target domains.
- Foreground (Things): Small, easily confused objects like pedestrians or signs are aligned using a triplet loss strategy. This pulls features of the same class together while pushing different classes apart in the feature space.
3. Adaptation Ability Estimation (AAE)
To further mitigate noise, the model calculates the variance between predictions from different branches. High variance indicates low confidence or disagreement, signaling an unreliable pseudo-label. The AAE module uses this variance to rectify the influence of these labels during training, effectively weighting the supervision by its estimated reliability.
Experimental Results
The framework was evaluated on standard "synthetic-to-real" pipelines:
- GTA5 → Cityscapes: EPS-UDA achieved a Mean Intersection over Union (mIoU) of 54.1%, outperforming previous state-of-the-art methods like DACS (52.1%) and IAST (52.2%).
- SYNTHIA → Cityscapes: The method reached an mIoU of 50.1% across 16 classes.
Ablation studies confirmed the importance of each component. While the tri-learning architecture provided a strong baseline, the addition of SFA (+4.4 mIoU) and AAE (+3.2 mIoU) was critical for achieving top-tier performance.
Conclusion
EPS-UDA effectively addresses the confirmation bias problem in UDA through its implicit pseudo-supervision mechanism. By leveraging the diversity of multiple classifiers and intelligent feature alignment, it bridges the domain gap between synthetic and real-world data more effectively than standard self-training. Future work may focus on reducing the computational overhead of training multiple segmentation heads and exploring more aggressive ways to maintain branch diversity.
Summary updated