Multi-Anchor Active Domain Adaptation for Semantic Segmentation

Abstract

Unsupervised domain adaption has been widely adopted in tasks with scarce annotated data. Unfortunately, mapping the target-domain distribution to the source-domain unconditionally may distort the essential structural information of the target-domain data, leading to inferior performance. To address this issue, we firstly propose to introduce active sample selection to assist domain adaptation regarding the semantic segmentation task. By innovatively adopting multiple anchors instead of a single centroid, both source and target domains can be better characterized as multimodal distributions, in which way more complementary and informative samples are selected from the target domain. With only a little workload to manually annotate these active samples, the distortion of the target-domain distribution can be effectively alleviated, achieving a large performance gain. In addition, a powerful semi-supervised domain adaptation strategy is proposed to alleviate the long-tail distribution problem and further improve the segmentation performance. Extensive experiments are conducted on public datasets, and the results demonstrate that the proposed approach outperforms stateof-the-art methods by large margins and achieves similar performance to the fully-supervised upperbound, i.e., 71.4% mIoU on GTA5 and 71.8% mIoU on SYNTHIA. The effectiveness of each component is also verified by thorough ablation studies. Code is available at https://github.com/munanning/MADAv2.

Unsupervised Domain Adaptation for Semantic Image Segmentation: a Comprehensive Survey

“…Related to semi-supervised DASiS is the active DASiS [138,170] where, instead of assuming that a small set of target samples are already labeled, an algorithm selects itself the images or pixels to be annotated by human annotators and use them to update the segmentation model over iterations -in a semi-supervised fashion. Ning et al [138] propose a multi-anchor based active learning strategy to identify the most complementary and representative samples for manual annotation by exploiting the feature distributions across the target and source domains. Shin et al [170] inspired by the maximization classifier discrepancy (MCD) for DA [160] -propose a method that selects the regions to be annotated based on the mismatch in predictions across the two classifiers.…”

Towards Fewer Annotations: Active Learning via Region Impurity and Prediction Uncertainty for Domain Adaptive Semantic Segmentation

“…To name a few, Prabhu et al [43] combine the uncertainty and diversity into an acquisition round and integrate semi-supervised domain adaptation into a unified framework. Lately, Ning et al [40] and Shin et al [53] are among the first to study the task of ADA applied to semantic segmentation, which greatly enhances the segmentation performance on the target domain. Ning et al [40] put forward a multi-anchor strategy to actively select a subset of images and annotate the entire image, which is probably inefficient.…”

“…Lately, Ning et al [40] and Shin et al [53] are among the first to study the task of ADA applied to semantic segmentation, which greatly enhances the segmentation performance on the target domain. Ning et al [40] put forward a multi-anchor strategy to actively select a subset of images and annotate the entire image, which is probably inefficient. While Shin et al [53] present a more efficient point-based annotation with an adaptive pixel selector.…”

“…To overcome this obstacle and encourage maximizing segmentation performance on the target domain, we show that a simple active learning strategy works well in adaptive semantic segmentation: annotating a small portion of im-age regions. Until recently, similar efforts have been made by Ning et al [40] and Shin et al [53]. The former uses multiple anchors to select representative target images to be labeled (Fig.…”

“…Illustration of different selection strategies. Imagebased selection (e.g., MADA [40]) picks a few target samples and label the entire image, which is probably inefficient. Point-based selection (e.g., LabOR [53]) chooses scarce points about which the model is uncertain, while uncertainty estimation at point level is prone to lump pixels that come from particular categories.…”