Semi-supervised Domain Adaptation with Instance Constraints

By Jeff Donahue, Judy Hoffman, Erik Rodner, Kate Saenko, Trevor Darrell. 2013

190 citations recorded by the source corpus.

Most successful object classification and detection methods rely on classifiers trained on large labeled datasets. However, for domains where labels are limited, simply borrowing labeled data from existing datasets can hurt performance, a phenomenon known as "dataset bias." We propose a general framework for adapting classifiers from "borrowed" data to the target domain using a combination of available labeled and unlabeled examples. Specifically, we show that imposing smoothness constraints on the classifier scores over the unlabeled data can lead to improved adaptation results. Such constraints are often available in the form of instance correspondences, e.g. when the same object or individual is observed simultaneously from multiple views, or tracked between video frames. In these cases, the object labels are unknown but can be constrained to be the same or similar. We propose techniques that build on existing domain adaptation methods by explicitly modeling these relationships, and demonstrate empirically that they improve recognition accuracy in two scenarios, multicategory image classification and object detection in video.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

LSDA: Large Scale Detection through Adaptation (2014, methodology): “These include variants for linear support vector machines [17, 18, 19], as well as adaptive latent SVMs [20] and adaptive exemplar SVM [21].”
Domain Adaptation for Visual Applications: A Comprehensive Survey (2017, methodology): “source templates [196, 197] with labeled target bounding boxes (SS scenario), and the adapted template is used at test time to detect the presence or absence of an object class in sliding windows. In [198] the PMT-SVM was further combined with MMDT [72] to handle complex domain shifts. The detector ”
Transfer Neural Trees for Heterogeneous Domain Adaptation (2016, methodology): “When it comes to the availability of labeled instances in the target domain, one can solve homogeneous DA under supervised [8], semi-supervised [9, 13], or unsupervised settings [10–12].”
Deep Unsupervised Domain Adaptation: A Review of Recent Advances and Perspectives (2022, methodology): “Further, several SSDA classification methods have been proposed to use instance constraints (Donahue et al., 2013), subspace learning (Yao et al.”
Reducing Domain Gap via Style-Agnostic Networks (2019, methodology): “It is commonly addressed by simultaneously minimizing the distance between the domains as well as imposing a regularization constraint on the target data to prevent overfitting [7,50].”
Automatic Adaptation of Object Detectors to New Domains Using Self-Training (2019, background): “and selecting target samples based on the baseline detector confidence score [19, 54, 47, 10, 30, 6].”
Visual and Semantic Knowledge Transfer for Large Scale Semi-Supervised Object Detection (2016, background): “general categories of TL have been proposed in previous work: homogeneous TL [13], [39], [40] in a single domain but with different data distributions in training and testing sets, and heterogeneous TL [41], [42], [43] across different domains or modalities.”
A Unified Framework for Metric Transfer Learning (2017, background): “As a third example, we consider the problem of object recognition [8], [9].”
Deep Transfer Network: Unsupervised Domain Adaptation (2015, methodology): “Such methods are being widely used in object classiﬁcation [1], object detection [6, 14] one-shot learning [9] etc .”
Fast Generalized Distillation for Semi-Supervised Domain Adaptation (2017, background): “…j=1 λj+1 ( y∗ij , ft(xi) ) ] s.t. ∑ i λi = 1 (3) Compared to other studies on SDA where each example of the source domain was used, by either re-weighting (Donahue et al. 2013; Duan et al. 2012) or augmentation (Daumé III, Kumar, and Saha 2010), GDSDA only requires the trained model from…”
Bayesian Joint Modelling for Object Localisation in Weakly Labelled Images (2015, background): “More recently, transductive SVM [37], Multiple kernel learning (MKL) [38], and instance constraints [39] have been exploited.”
DOCK: Detecting Objects by Transferring Common-Sense Knowledge (2018, background): “Existing transfer learning approaches can be roughly divided into two groups: one that learns using bounding box annotations for both source and target categories [36,3,21,11,43,40], and another that learns using bounding box annotations for source categories but only image-level annotations for tar”
Reducing Domain Gap by Reducing Style Bias (2021, background): “It is commonly addressed by simultaneously minimizing the distance between the domains as well as imposing a regularization constraint on the target data to prevent overﬁt-ting [7, 50].”
Expanding object detector's Horizon: Incremental learning framework for object detection in videos (2015, methodology): “Perhaps closest works to ours are [31] and [6].”
Weakly-Shared Deep Transfer Networks for Heterogeneous-Domain Knowledge Propagation (2015, background): “Transfer learning can be mainly categorized two subsets: homogeneous transfer learning (domain adaptation) [7, 18, 21, 14] in a single domain but with different distributions in training and testing sets, and heterogenous transfer learning [37, 8, 24] across different modalities.”
Semi-Supervised Domain Adaptation via Minimax Entropy (2019, background): “Semi-supervised domain adaptation (SSDA) is a very important task [8, 40, 1], however it has not been fully explored, especially with regard to deep learning based methods.”
Heterogeneous Domain Adaptation for Multiple Classes (2014, background): “where λi > 0 is the regularization parameter, and the nonnegativity constraints are to preserve nonnegative correlation of the source and target weight vectors [3, 12].”