Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation
Authors: Hung-Chieh Fang, Po-Yi Lu, Hsuan-Tien Lin
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results confirm that SSL consistently advances PDM and delivers new stateof-the-art results across a broader benchmark of Uni DA scenarios with different portions of shared classes, representing a crucial step toward truly comprehensive Uni DA. Project page: https://dc-unida.github.io/ |
| Researcher Affiliation | Academia | 1National Taiwan University. Correspondence to: Hung Chieh Fang <EMAIL>, Hsuan-Tien Lin <EMAIL>. |
| Pseudocode | No | The paper describes methods and equations, such as Ls(θf, θc) and Ladv(θf, θd), but does not present structured pseudocode or algorithm blocks. |
| Open Source Code | No | Project page: https://dc-unida.github.io/ |
| Open Datasets | Yes | We present results on four widely used benchmarks: Office31, Office Home, Vis DA, and Domain Net. Details of these datasets can be found in Appendix D.3. Office31 (Saenko et al., 2010) contains 31 classes and three domains: Amazon (A), DSLR (D), and Webcam (W), with a total of about 4k images. Office-Home (Venkateswara et al., 2017) has 65 classes and four domains: Art (A), Product (Pr), Clipart (Cl), and Realworld (Rw), with approximately 15k images. Vis DA (Peng et al., 2017) is a larger dataset with 12 classes from two domains: Synthetic and Real images, totaling around 280k images. Domain Net (Peng et al., 2019), the largest DA dataset, has 345 classes and six domains, with about 0.6 million images. |
| Dataset Splits | No | The paper describes how classes are divided into source-private, common, and target-private sets (e.g., Table 10) and refers to the 'conventional setup used in prior works' for general Uni DA settings. However, it does not specify the train/test/validation splits (percentages or exact counts) for the datasets used in the experiments. |
| Hardware Specification | No | We thank to National Center for High-performance Computing (NCHC) of National Applied Research Laboratories (NARLabs) in Taiwan for providing computational and storage resources. |
| Software Dependencies | No | The paper mentions using ResNet-50 as the backbone model and Sim Siam as the self-supervised loss, but it does not list specific version numbers for software libraries or dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The training steps are 10K for all experiments and the batch size is set to 36 for both domains. The hyperparameters are set as follows: λAdv = 0.5 and λSSL = 0.5 for Office-Home, Domain Net and Vis DA, and λSSL = 0.2 for Office31. We use Sim Siam (Chen & He, 2021) as our self-supervised loss as it does not require negative samples or large batch size. The data augmentation strategy follows the same setup as Sim Siam. |