Related Experiment Video
Updated: Jun 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Reducing semantic ambiguity in domain adaptive semantic segmentation via probabilistic prototypical pixel contrast
Xiaoke Hao1, Shiyu Liu1, Chuanbo Feng1
1School of Artificial Intelligence, Hebei University of Technology, No. 5340, Xiping Road, Tianjin, 300401, China.
Summary
This study introduces Probabilistic Prototypical Pixel Contrast (PPPC), a novel domain adaptation framework. PPPC enhances model performance by modeling pixel embeddings as probabilities, effectively addressing ambiguity in visual data.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Domain adaptation techniques aim to mitigate performance degradation caused by domain shifts in machine learning models.
- Existing methods combining contrastive learning and self-training struggle with ambiguous scenarios due to deterministic pixel embeddings.
Purpose of the Study:
- To propose a universal domain adaptation framework, Probabilistic Prototypical Pixel Contrast (PPPC), that improves representation quality by addressing pixel-level ambiguity.
- To enhance the robustness and generalization of models in challenging domain adaptation tasks.
Main Methods:
- PPPC models each pixel embedding as a probability distribution using a multivariate Gaussian distribution, capturing inherent uncertainty.
- Prototypes are derived from posterior probability estimation to refine decision boundaries, pushing them away from ambiguous regions.
- An efficient method for computing distribution similarity is employed, avoiding sampling and reparameterization to reduce computational overhead.
- Ambiguous image crops are dynamically selected to increase boundary point involvement in contrastive learning, leading to more precise category distributions.
Main Results:
- PPPC effectively addresses pixel-level ambiguity, leading to more discriminative representations.
- Significant improvements were observed in both synthetic-to-real and day-to-night domain adaptation tasks.
- The framework achieved a +5.2% mIoU improvement over state-of-the-art methods in challenging daytime-to-nighttime adaptation scenarios.
- Demonstrated stronger generalization capabilities on unseen datasets.
Conclusions:
- PPPC offers a robust and effective solution for domain adaptation by leveraging probabilistic modeling of pixel embeddings.
- The proposed method significantly enhances model performance and generalization, particularly in scenarios with significant domain shifts and visual ambiguities.
- The framework provides a promising direction for future research in domain adaptation and representation learning.
