Domain Adaptation

A vision model trained in one setting can look impressively accurate, then fail when the setting changes. A pedestrian detector trained on sunny city photos, for example, may struggle with night-time dashcam footage because the images have different lighting, noise, camera angles, and backgrounds. Domain adaptation is the set of techniques used to help a model carry its knowledge from a familiar visual environment into a different one.

What is being adapted?
The original training environment is the source domain; the new environment is the target domain. The task usually stays the same—such as identifying cars, reading text, or segmenting tumors—but the image distribution changes. Domain adaptation tries to reduce this domain shift without requiring a large, fully labeled target dataset. Common settings include:

• Unsupervised domain adaptation: labeled source images and unlabeled target images are available.
• Supervised domain adaptation: a small number of target images have labels.
• Source-free adaptation: only a trained source model is available; its original training data cannot be shared.

How models bridge the gap
A simple approach is to fine-tune a pretrained model on a small labeled target set. More specialized methods teach the model to extract features that work for both domains. For instance, Domain-Adversarial Neural Networks (DANN) add a second classifier that tries to tell whether a feature came from the source or target domain. The feature extractor is trained to defeat that classifier while still solving the vision task. The result is a representation that preserves useful signals, such as object shape, while becoming less dependent on camera-specific appearance.

Why it matters in practice
Domain adaptation makes vision systems more reliable when collecting labels is expensive or impractical. It supports transferring synthetic driving scenes to real roads, adapting factory defect detectors between production lines, and moving medical segmentation models between hospitals with different scanners. Without it, a model can confuse harmless lighting or device differences with meaningful visual evidence—an especially costly mistake in safety and quality-critical systems.