VisDA: The Visual Domain Adaptation Challenge

By Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, Kate Saenko. 2017

929 citations recorded by the source corpus.

We present the 2017 Visual Domain Adaptation (VisDA) dataset and challenge, a large-scale testbed for unsupervised domain adaptation across visual domains. Unsupervised domain adaptation aims to solve the real-world problem of domain shift, where machine learning models trained on one domain must be transferred and adapted to a novel visual domain without additional supervision. The VisDA2017 challenge is focused on the simulation-to-reality shift and has two associated tasks: image classification and image segmentation. The goal in both tracks is to first train a model on simulated, synthetic data in the source domain and then adapt it to perform well on real image data in the unlabeled test domain. Our dataset is the largest one to date for cross-domain object classification, with over 280K images across 12 categories in the combined training, validation and testing domains. The image segmentation dataset is also large-scale with over 30K images across 18 categories in the three domains. We compare VisDA to existing cross-domain adaptation datasets and provide a baseline performance analysis using various domain adaptation models that are currently popular in the field.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

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

In Representation & Self-Supervised Learning

Citations in the corpus

Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100 (2020, methodology): “wo years later’. We further split target videos into: Target Train and Target Test. The rst are unlabelled videos used during domain adaptation, while the second are videos used for evaluation, as in [88]. Number of action instances per split are in Table 8. Evaluation. We use the same evaluation me”
Moment Matching for Multi-Source Domain Adaptation (2018, background): “More recent benchmarks [43, 17, 34] are proposed to evaluate the effectiveness of domain adaptation models.”
Class-Incremental Domain Adaptation (2020, methodology): “In this section, we provide the details of the architectures used for Office [11], VisDA [9] and Digits datasets, along with the architectures of the baseline models presented in the paper (§4.”
Contrastive Adaptation Network for Unsupervised Domain Adaptation (2019, methodology): “We validate our method on two public UDA benchmarks: Ofﬁce-31 [25] and VisDA-2017 [24].”
Generate to Adapt: Aligning Domains Using Generative Adversarial Networks (2017, methodology): “In this section, we present the results on VISDA dataset [26] - a large scale testbed for unsupervised domain adaptation algorithms.”
TTT++: When Does Self-Supervised Test-Time Training Fail or Thrive? (2021, methodology): “We ﬁnally validate the effectiveness of our method on the VisDA-C dataset [42], a challenging large-scale benchmark of synthetic-to-real object classiﬁcation.”
Open Set Domain Adaptation by Backpropagation (2018, methodology): “We conduct experiments on Oﬃce [3], VisDA [28] and digits datasets.”
Denoised Maximum Classifier Discrepancy for Source-Free Unsupervised Domain Adaptation (2022, methodology): “For the VISDA dataset, the learning rates for the feature extractor and the feature classifier are set to 3 × 10 − 4 and 1 × 10 − 3 respectively.”
Sim-Real Joint Reinforcement Transfer for 3D Indoor Navigation (2019, background): “There are several work [16, 41] focus on Synthetic-to-Real visual image translation and a benchmark called VisDA [28] has been proposed recently.”
Improving Domain-Specific Classification by Collaborative Learning with Adaptation Networks (2019, methodology): “2011), the object datasets: CIFAR-10 (Krizhevsky and Hinton 2009) and STL-10 (Coates, Lee, and Ng 2011), and the VisDA-2017 dataset (Peng et al. 2017).”
Is synthetic data from generative models ready for image recognition? (2022, background): “The ﬁrst type, synthetic datasets (Dosovitskiy et al., 2015; Peng et al., 2017; Richter et al., 2016), are usually generated from a traditional pipeline with a speciﬁc data source, e . g .synthetic 2D renderings of 3D models or scenes from graphics engines.”
Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation (2021, background): “VisDA [28] is a more challenging dataset, with 12-class synthetic-to-real object recognition tasks, its source domain contains of 152k synthetic images while the target domain has 55k real object images.”
Characterizing and Avoiding Negative Transfer (2018, methodology): “ we consider three transfer tasks: Ar!Rw, Cl!Rw, and Pr!Rw. In addition, we choose to use the ﬁrst 25 categories in alphabetic order to make our results more comparable to previous studies [4]. VisDA [22] is another challenging synthetic to real dataset. We use the training set as the synthetic sour”
Enhancing Photorealism Enhancement (2021, background): “Prior work hypothesized that differences in scene layout may be a key contributing factor in suboptimal performance in image translation [19] or in training semantic segmentation models [81], [82], and addressed this challenge via regularization [78], semantic consistency losses [19], [83], and matc”
Maximum Classifier Discrepancy for Unsupervised Domain Adaptation (2017, methodology): “The VisDA dataset [28] was used in this experiment, which evaluated adaptation from synthetic-object to real-object images.”
Adversarial Dropout Regularization (2017, methodology): “We use a new domain adaptation benchmark called the VisDA Challenge (Peng et al. (2017)) which focuses on the challenging task of adapting from synthetic to real images.”
Fully Convolutional Adaptation Networks for Semantic Segmentation (2018, methodology): “Following the settings in [9, 21], 1,500 frames are used for evaluation.”
Domain2Vec: Domain Embedding for Unsupervised Domain Adaptation (2020, methodology): “Instead, in our work, we demo our model on the following four transfer setting: DomainNet [16] → VisDA [53], DomainNet → Youtube BBox [72], DomainNet → PASCAL [64], DomainNet → COCO.”