Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation
Authors: Shiqi Yang, yaxing wang, Joost van de Weijer, Luis Herranz, Shangling Jui
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experimental results we verify that the inherent structure of the target features is an important source of information for domain adaptation. We demonstrate that this local structure can be efficiently captured by considering the local neighbors, the reciprocal neighbors, and the expanded neighborhood. Finally, we achieve state-of-the-art performance on several 2D image and 3D point cloud recognition datasets. Code is available in https://github.com/Albert0147/SFDA_neighbors. 4 Experiments Datasets. We use three 2D image benchmark datasets and a 3D point cloud recognition dataset. Office-31 [32] contains 3 domains (Amazon, Webcam, DSLR) with 31 classes and 4,652 images. Table 1: Accuracies (%) on Office-31 for Res Net50-based methods. |
| Researcher Affiliation | Collaboration | Shiqi Yang1, Yaxing Wang1,2 , Joost van de Weijer1, Luis Herranz1, Shangling Jui3 1 Computer Vision Center, Universitat Autonoma de Barcelona, Barcelona, Spain 2 PCALab, Nanjing University of Science and Technology, China 3 Huawei Kirin Solution, Shanghai, China {syang,yaxing,joost,lherranz}@cvc.uab.es, jui.shangling@huawei.com |
| Pseudocode | Yes | Algorithm 1 Neighborhood Reciprocity Clustering for Source-free Domain Adaptation |
| Open Source Code | Yes | Code is available in https://github.com/Albert0147/SFDA_neighbors. |
| Open Datasets | Yes | Datasets. We use three 2D image benchmark datasets and a 3D point cloud recognition dataset. Office-31 [32] contains 3 domains (Amazon, Webcam, DSLR) with 31 classes and 4,652 images. Office-Home [46] contains 4 domains (Real, Clipart, Art, Product) with 65 classes and a total of 15,500 images. Vis DA [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. Point DA-10 [30] is the first 3D point cloud benchmark specifically designed for domain adaptation, it has 3 domains with 10 classes, denoted as Model Net-10, Shape Net-10 and Scan Net-10, containing approximately 27.7k training and 5.1k testing images together. |
| Dataset Splits | Yes | Datasets. We use three 2D image benchmark datasets and a 3D point cloud recognition dataset. Office-31 [32] contains 3 domains (Amazon, Webcam, DSLR) with 31 classes and 4,652 images. Office-Home [46] contains 4 domains (Real, Clipart, Art, Product) with 65 classes and a total of 15,500 images. Vis DA [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. Point DA-10 [30] is the first 3D point cloud benchmark specifically designed for domain adaptation, it has 3 domains with 10 classes, denoted as Model Net-10, Shape Net-10 and Scan Net-10, containing approximately 27.7k training and 5.1k testing images together. |
| Hardware Specification | Yes | Experiments are conducted on a TITAN Xp. |
| Software Dependencies | No | The paper does not explicitly state specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or other libraries). |
| Experiment Setup | Yes | We adopt SGD with momentum 0.9 and batch size of 64 for all 2D datasets, and Adam for Point DA-10. The learning rate for Office-31 and Office-Home is set to 1e-3 for all layers, except for the last two newly added fc layers, where we apply 1e-2. Learning rates are set 10 times smaller for Vis DA. Learning rate for Point DA-10 is set to 1e-6. We train 30 epochs for Office-31 and Office Home while 15 epochs for Vis DA, and 100 for Point DA-10. For the number of nearest neighbors (K) and expanded neighborhoods (M), we use 3,2 for Office-31, Office-Home and Point DA-10, since Vis DA is much larger we set K, M to 5. |