1. Debasmit Das and C. S. George Lee, Purdue University, West Lafayette, IN, USA {dsdas, csglee}@purdue.edu
Unsupervised Domain Adaptation
Using Regularized Hyper-graph Matching
1. INTRODUCTION
Training and Testing
Distributions might
be different
2. RELATED WORK
3. APPROACH
4. FIND EXEMPLARS
5. FIND CORRESPONDENCE
8. ACKNOWLEDGEMENT
7. EXPERIMENTS
This work was supported in part by the National Science Foundation
under Grant IIS-1813935. Any opinion, findings, and conclusions or
recommendations expressed in this material are those of the authors
and do not necessarily reflect the views of the National Science
Foundation.
Find Exemplars from
both Domains
Find Matching
between exemplar
hyper-graphs
Training Samples Testing Samples
Categories same across domains
Target domain is unlabeled
Map source domain
to target domain
Dataset : Office-Caltech
4 domains : Amazon (A),
Caltech (C), Webcam
(W), DSLR (D)
Feature : SURF
Discrepancy Based Methods
Mostly global metrics. Minimize statistics of data like covariance
[Sun et al. ECCV’16] or maximum mean discrepancy [Long et al. ICML’15]
Local Method
Optimal Transport. Only First
Order information used.
[Courty et al. TPAMI’17]
2nd order matching1st order matching
3rd order matching
Group-Lasso
6. SOLUTION
Consensus ADMMCombine Conditional
gradient and ADMM
Update TC linear
compared to Interior
Point’s cubic
• Extract Exemplars using affinity
propagation
• Use Similarity Matrix between data points
• Message passing between data points Computational
efficiency
Avoid
Outliers