Factors of Influence for Transfer Learning Across Diverse Appearance Domains and Task Types.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Transfer learning reuses knowledge from source tasks for target tasks.
• Current computer vision models often use ImageNet (ILSVRC) pre-training, but its effectiveness across diverse domains is not fully understood.

Purpose of the Study:

• To systematically explore transfer learning across various image domains and complex computer vision tasks.
• To identify factors influencing transfer learning performance and provide concrete recommendations.

Main Methods:

• Conducted over 2000 transfer learning experiments.
• Investigated transfer across diverse image domains (e.g., consumer photos, autonomous driving, aerial) and task types (e.g., segmentation, detection, depth estimation).
• Analyzed the impact of image domain, task type, and dataset size on performance.

Main Results:

• Found sources other than ILSVRC'12 often outperform it.
• Image domain similarity is the most critical factor for positive transfer.
• Source datasets including the target domain yield best results.
• Transfer across task types is beneficial but task-dependent.

Conclusions:

• Image domain alignment is crucial for successful transfer learning.
• Recommendations are provided for selecting optimal source datasets and tasks for improved transfer learning performance in computer vision.