Learning from Synthetic Images via Active Pseudo-Labeling.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Synthetic visual data offers cost-effective, accurately annotated datasets.
• Domain gaps in appearance and label distribution limit the direct application of models trained on synthetic data to real-world scenarios.
• Manual annotation of real-world data is labor-intensive and time-consuming.

Purpose of the Study:

• To develop a novel framework, Active Pseudo-Labeling (APL), to reduce domain gaps between synthetic and real visual data.
• To enable effective learning from synthetic data for real-world computer vision tasks.
• To improve the performance of models on real-world datasets without extensive manual labeling.

Main Methods:

• The proposed Active Pseudo-Labeling (APL) framework adapts the style of real images to the synthetic source domain using a task-guided generative model.
• Pseudo-labels are predicted for these style-adapted real images.
• A model pre-trained on synthetic data is fine-tuned on the pseudo-labeled real data to align with the target domain distribution.

Main Results:

• Experiments on semantic segmentation and object detection tasks demonstrated the effectiveness of the APL framework.
• The proposed method showed superior performance compared to existing state-of-the-art approaches in reducing domain gaps.
• APL successfully improved model generalization from synthetic to real visual data.

Conclusions:

• Active Pseudo-Labeling (APL) is a significant advancement in leveraging synthetic data for real-world computer vision applications.
• The framework effectively bridges the domain gap, enhancing model performance on real datasets.
• APL offers a promising solution for scenarios where real-world data annotation is challenging.