Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration

Abstract

Source-free domain adaptation (SFDA) aims to adapt a model trained on labelled data in a source domain to unlabelled data in a target domain without access to the source-domain data during adaptation. Existing methods for SFDA leverage entropy-minimization techniques which: (i) apply only to classification; (ii) destroy model calibration; and (iii) rely on the source model achieving a good level of feature-space class-separation in the target domain. We address these issues for a particularly pervasive type of domain shift called measurement shift, characterized by a change in measurement system (e.g. a change in sensor or lighting). In the source domain, we store a lightweight and flexible approximation of the feature distribution under the source data. In the target domain, we adapt the featureextractor such that the approximate feature distribution under the target data realigns with that saved on the source. We call this method Feature Restoration (FR) as it seeks to extract features with the same semantics from the target domain as were previously extracted from the source. We additionally propose Bottom-Up Feature Restoration (BUFR)-a bottom-up training scheme for FR which boosts performance by preserving learnt structure in the later layers of a network. Through experiments we demonstrate that BUFR often outperforms existing SFDA methods in terms of accuracy, calibration, and data efficiency, while being less reliant on the performance of the source model in the target domain.

Real-Time Style Modelling of Human Locomotion via Feature-Wise Transformations and Local Motion Phases

“…For example, since the development of a neural approach by Gatys et al [17], many methods have been designed that apply transformations to hidden units of convolutional networks to stylise images [12,27,28]. Style transfer is also closely related to the wider task of domain adaptation [4,13,14,25,45,52].…”

Towards Data-Free Domain Generalization

“…SHOT [40] employs an information maximization loss along with a self-supervised pseudo-labeling, and is extended to the multi-source scenario via source model weighting [1]. BUFR [13] aligns the target domain feature distribution with the one from the source domain. Another line of works leverage Batch Normalization (BN) [28] layers by replacing the BN-statistics computed on the source domain with those computed on the target domain [38], or by training the BN-parameters on the target domain via entropy minimization [74].…”

“…While these approaches rely on the availability of data from a known target domain, we address the DFDG scenario where the model is expected to generalize to a priori unknown target domain(s) without any modification or exposure to their data. We also note that some methods [31,40,13] modify the training procedure on the source domain, which would not be possible in cases where the data is not accessible anymore.…”

“…Although Data-Free Knowledge Distillation (DFKD) methods were developed to transfer knowledge from a teacher model to a student model without any access to the original data [42,47,8,53,82,10], only single-teacher scenarios with no domain shift were studied. On the other hand, Source-Free Domain Adaptation (SFDA) approaches were proposed to tackle the domain shift problem setting where one [40,36,31,74,13] or multiple [1] models trained on source domain data are available instead of the original dataset(s). Nonetheless, they require access to data from the target domain.…”