Unsupervised Domain Adaptation of Object Detectors: A Survey

Abstract

Recent advances in deep learning have led to the development of accurate and efficient models for various computer vision applications such as object classification, semantic segmentation, and object detection. However, learning highly accurate models relies on the availability of datasets with large number of annotated images. Due to this, model performance drops drastically when evaluated on label-scarce datasets having visually distinct images. This issue is commonly referred to as covariate shift or dataset bias. Domain adaptation attempts to address this problem by leveraging domain shift characteristics from labeled data in a related domain when learning a classifier for label-scarce target dataset. There are a plethora of works to adapt object classification and semantic segmentation models to label-scarce target dataset through unsupervised domain adaptation. Considering that object detection is a fundamental task in computer vision, many recent works have recently focused on addressing the domain adaptation issue for object detection as well. In this paper, we provide a brief introduction to the domain adaptation problem for object detection and present an overview of various methods proposed till date for addressing this problem. Furthermore, we highlight strategies proposed for this problem and the associated shortcomings. Subsequently, we identify multiple aspects of the unsupervised domain adaptive detection problem that are most promising for the future research in the area. We believe that this survey shall be valuable to the pattern recognition experts working in the fields of computer vision, biometrics, medical imaging, and autonomous navigation by introducing them to the problem, getting them familiar with the current status of the progress and providing with promising direction for future research.

Unsupervised Domain Adaptation: A Reality Check

“…Tables [20][21][22][23] show the performance gap between oracle and non-oracle validators, per algorithm, at a 0.98 source threshold. Dashes indicate that either all models were discarded with the 0.98 threshold, or that those algorithm/validator/task combinations had not yet run.…”

“…Applications of UDA include semantic segmentation [41], object detection [21], and natural language processing [26]. There are also other types of domain adaptation, including semi-supervised [30], multi-source [24], partial [2,22,34], universal [49], and source-free [14].…”

Adversarially Trained Object Detector for Unsupervised Domain Adaptation

“…Many domain adaptation approaches have also been proposed for object detection (Oza et al 2021). Typical approaches include adversarial feature learning (Chen et al 2018;Saito et al 2019;Chen et al 2021b), pseudo-labelbased self-training (Kim et al 2019;Li et al 2021), and image-to-image translation (Hsu et al 2020;Chen et al 2020).…”

“…main adaptation assumes that the target domain has no labels (Zhao et al 2020). In recent studies, several approaches have been proposed for implementing unsupervised domain adaptation in object detection tasks (Oza et al 2021). The most common approach is adversarial feature learning, which involves aligning the source and target features using a feature extractor competing with a domain discriminator (Chen et al 2018;Saito et al 2019;Chen et al 2021b).…”

Seeing BDD100K in dark: Single-Stage Night-time Object Detection via Continual Fourier Contrastive Learning

“…However, as the images from daytime and night-time vary largely with respect to their input domain properties, a detector trained on source domain daytime labeled data alone might not perform well on target domain night-time images. To bridge this domain discrepancy, one may leverage additional unlabeled nighttime images along with labeled daytime images, leading to an Unsupervised Domain Adaptation (DA) setting [28,33].…”