Label Calibration in Source Free Domain Adaptation

TL;DR

This work tackles source-free domain adaptation by refining noisy target pseudolabels through an evidential learning framework that places a Dirichlet prior over target predictions to quantify distributional uncertainty, paired with calibrated softmax to address translation invariance. The method comes in two variants: EKS (with prior knowledge) and ES (without prior knowledge), and combines EDl losses ($L_{edl} = L_{nll} + \beta L_{kl}$) with calibrated information maximization ($\hat{L}_{im} = \hat{L}_{ent} + \hat{L}_{div}$) to form a robust training objective $L_{total} = w_1 L_{edl} + w_2 \hat{L}_{im}$. Prior knowledge is encoded via unary bounds and binary relationships on target class probabilities, solvable with a solver, while calibrated softmax ($\hat{\delta}$) reduces overconfidence and improves diversity in predictions. Extensive experiments on Domainnet40, Office-Home, Office31, and Digits demonstrate consistent improvements over SHOT/KSHOT baselines, with notable gains when prior knowledge is available and under large domain shifts. The approach yields better calibration (lower ECE, favorable NLL) and clearer, more discriminative feature representations, as shown by visualization analyses.

Abstract

Source-free domain adaptation (SFDA) utilizes a pre-trained source model with unlabeled target data. Self-supervised SFDA techniques generate pseudolabels from the pre-trained source model, but these pseudolabels often contain noise due to domain discrepancies between the source and target domains. Traditional self-supervised SFDA techniques rely on deterministic model predictions using the softmax function, leading to unreliable pseudolabels. In this work, we propose to introduce predictive uncertainty and softmax calibration for pseudolabel refinement using evidential deep learning. The Dirichlet prior is placed over the output of the target network to capture uncertainty using evidence with a single forward pass. Furthermore, softmax calibration solves the translation invariance problem to assist in learning with noisy labels. We incorporate a combination of evidential deep learning loss and information maximization loss with calibrated softmax in both prior and non-prior target knowledge SFDA settings. Extensive experimental analysis shows that our method outperforms other state-of-the-art methods on benchmark datasets.

Figures (3)

• Figure 1: Overview of the proposed EKS method. The dirichlet distribution and calibrated softmax is placed over logits of the target model followed by adaptation with EDL loss and IM loss in prior knowledge SFDA setting. In the proposed ES method, there is non-availability of prior knowledge, and self-supervised pseudolabels are directly used for EDL loss (best view in color).
• Figure 2: t-SNE plots for P$\rightarrow$C and S$\rightarrow$C on Domainnet40.
• Figure 3: Expected Calibration Error (ECE) curve on the Domainnet40 dataset.