Distribution-Matching Embedding for Visual Domain Adaptation
Authors: Mahsa Baktashmotlagh, Mehrtash Harandi, Mathieu Salzmann
JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the benefits of our approach on the tasks of visual object recognition, text categorization, and Wi Fi localization. Our experiments have evidenced the importance of exploiting distribution invariance for domain adaptation by revealing that our algorithms yield state-of-the-art results on several problems. 6. Experiments We evaluated our approach on the tasks of visual object recognition, cross-domain text categorization, and cross-domain Wi Fi localization, and compare its performance against the state-of-the art methods in each task. |
| Researcher Affiliation | Collaboration | Mahsa Baktashmotlagh EMAIL Queensland University of Technology Brisbane, Australia Mehrtash Harandi EMAIL Australian National University & NICTA Canberra, Australia Mathieu Salzmann EMAIL CVLab, EPFL Lausanne, Switzerland. NICTA is funded by the Australian Government as represented by the Department of Broadband, Communications and the Digital Economy and the ARC through the ICT Centre of Excellence program. |
| Pseudocode | Yes | Algorithm 1 : Learning on a Grassmann Manifold |
| Open Source Code | No | The paper provides a link to code for a baseline method (GFK) used for comparison, stating: "The code can be downloaded from: http://users.cecs.anu.edu.au/ basura/DA_SA/get GFKDim.m". However, it does not explicitly state that the authors' own implementation code for the Distribution-Matching Embedding (DME) or Nonlinear DME (NL-DME) methods is available or provide a link to it. |
| Open Datasets | Yes | We evaluated our approach on the tasks of visual object recognition using the benchmark domain adaptation data set introduced in (Saenko et al., 2010)... we made use of the 20 Newsgroups data set... we used the Wi Fi data set published in the 2007 IEEE ICDM Contest for domain adaptation (Yang et al., 2008). |
| Dataset Splits | Yes | For each source/target pair, we report the average recognition accuracy and standard deviation over the 20 partitions provided with GFK1... For this experiment, we used the protocol of (Duan et al., 2012): The four largest main categories (comp, rec, sci, and talk) were chosen for evaluation. Specifically, for each main category, the largest subcategory was selected as the target domain... To construct the training set, we used 1000 randomly selected samples (evenly distributed positive and negative samples) from the source domain, and repeated this procedure 5 times... The data set contains 621 labeled examples collected during time period A (i.e., the source) and 3128 unlabeled examples collected during time period B (i.e., the target). We followed the transductive setting of Pan et al. (2011), which uses all the samples from the source and 400 random samples from the target. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory, or cloud computing instance types used for running its experiments. |
| Software Dependencies | No | The paper mentions using "a kernel SVM classifier with a degree 2 polynomial kernel, or a linear SVM classifier" and "conjugate gradient (CG) algorithm on the Grassmann manifold," but does not specify any software names with version numbers for their implementation. |
| Experiment Setup | Yes | The only parameter of such classifiers is the regularizer weight C. For each method, we tested with C {10 5, 10 3, 10 2, 10 1, 100, 101, 102}, and report the best result... The bandwidth of the Gaussian RBF kernel used in MMD was taken as the median distance computed over all pairwise data points; the value σs, such that Hs = σs I in Eq. 15, was computed using the maximal smoothing principle (Terrell, 1990)... In practice, we initialize W to the truncated identity matrix... For all the methods based on dimensionality reduction, we used the dimensionalities provided in the GFK code (i.e., W-D: 10, D-A: 20, W-A: 10, C-W: 20, C-D: 10, C-A: 20)... For all the baselines based on dimensionality reduction, we set the dimensionalities based on the Subspace Disagreement Measure (SDM) (Gong et al., 2012)2 (i.e., comp vs. rec: 10, comp vs. sci: 27, comp vs. talk: 47)... Note that, here, all results were obtained with a nearest-neighbor classifier to follow the procedure of (Pan et al., 2011). |