Unsupervised Domain Adaptation for Visual Recognition

Module: Computer Vision | Difficulty: Premium

Domain Discrepancy

CORAL Loss

where are covariance matrices of source and target features.

DANN (Domain-Adversarial Training)

where GRL is the gradient reversal layer.

import torch
import torch.nn as nn
import torch.nn.functional as F

class GradientReversalLayer(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x, alpha):
        ctx.alpha = alpha
        return x.view_as(x)

    @staticmethod
    def backward(ctx, grad_output):
        return -ctx.alpha * grad_output, None

class DomainAdversarialNetwork(nn.Module):
    def __init__(self, feature_extractor, classifier,
                 domain_discriminator, alpha=1.0):
        super().__init__()
        self.feature_extractor = feature_extractor
        self.classifier = classifier
        self.domain_discriminator = domain_discriminator
        self.alpha = alpha

    def forward(self, x, return_domain=False):
        features = self.feature_extractor(x)
        class_output = self.classifier(features)
        reversed_features = GradientReversalLayer.apply(
            features, self.alpha)
        domain_output = self.domain_discriminator(reversed_features)
        if return_domain:
            return class_output, domain_output
        return class_output

class SelfTrainingDA:
    def __init__(self, model, threshold=0.95):
        self.model = model
        self.threshold = threshold

    def generate_pseudo_labels(self, target_data):
        self.model.eval()
        pseudo_labels = []
        with torch.no_grad():
            logits = self.model(target_data)
            probs = F.softmax(logits, dim=1)
            max_probs, labels = probs.max(dim=1)
            mask = max_probs > self.threshold
            pseudo_labels.append((target_data[mask], labels[mask]))
        return pseudo_labels

Research Insight: Domain adaptation has evolved from adversarial methods (DANN) to self-training approaches (DPL, DAFormer) that use pseudo-labeling on target data. The key insight is that high-confidence pseudo-labels provide effective supervision for adapting to the target domain. DAFormer demonstrated that transformer-based architectures are more robust to domain shift than CNNs, likely due to their global receptive field and data efficiency. The combination of masked image modeling with domain adaptation achieves state-of-the-art results by learning domain-invariant features.