Domain Adaptation in the Absence of Source Domain Data

Uncertainty Reduction for Model Adaptation in Semantic Segmentation

“…Model Adaptation Most of the previously mentioned methods need the explicit availability of source domain data during adaptation too, and have made tremendous strides in improving the segmentation performance in that case. A few recent papers tackle model adaptation for classification problems [44,42,16]. [39] proposes source-free domain adaptation in the case where label knowledge of the target domain is not available, and show their efficiency on a set of classification problems with varying levels of label overlap.…”

“…These pseudo-tasks train the shared network structure to learn the task in multiple ways, and can be interpreted as using self-supervision. Similarly, increasing the robustness of classification through feature noising has been studied extensively [28,8,68,16,47]. Work on learning robust networks through pseudoensembling by reducing the variance when dropout is used, has been proposed [2].…”

“…A large portion of the methods that has been devoted to tackling this problem, like some of the ones reviewed in Section 4, require the labeled source data to be available along with the unlabeled target data for the adaptation process. We, in this paper, focus on the problem where the source data itself is not available but the source trained classifier is [16]. This is similar to life-long learning [63], where the goal is to adapt to several tasks over several domains, and the only information payload that is carried over is the model itself.…”

“…Another relevant application where we envisage such a setting is in search-and-rescue operations, where data is collected on a mobile drone, and the segmentor network is adapted based on only the data collected, without having the need to access the original labeled dataset. Thus the problem of domain adaptation in the absence of source data, termed model adaptation [16], is of practical significance.…”

“…In the absence of source labeled data that has been effectively exploited by previous works, we enforce auxiliary properties that are desirable in a system, namely confident predictions for the target data, and noise resilience, and thereby increased stability of classification to parameter choices. To this end, we propose a method that uses feature corruption [16,47,54], and entropy regularization [29,69]. We find that having access only to the source classifier, along with unlabeled target data, can result in performance comparable to the case where source data is also available.…”