Source-free Domain Adaptation
Xin Zhao
CVML Lab@SMU Group Seminar
25 February 2022
2
Unsupervised Domain Adaptation (UDA)
Transfer the knowledge from the labeled source domain to an unlabeled target domain to learn an accurate model for target data.
Domain Shift
3
Unsupervised Domain Adaptation (UDA)
Conventional UDA methods normally assume that the source data is available during training.
4
Source-free Domain Adaptation (SFDA)
5
Two Papers
6
SHOT: Source Model Generation
7
SHOT: Motivation
8
SHOT: Source Hypothesis Transfer
9
SHOT: Source Hypothesis Transfer
10
SHOT: Source Hypothesis Transfer
11
SHOT: Source Hypothesis Transfer
12
SHOT: Source Hypothesis Transfer
13
SHOT: Classification Results
very helpful.
14
SHOT: Analysis
15
SHOT: Analysis
16
Two Papers
17
HCL
source-domain hypothesis.
18
HCL: Historical Contrastive Instance Discrimination (HCID)
current (as queries) and historical models (as keys): the positive pairs are pulled close while negative pairs are
pushed apart.
19
and historical models.
HCL: Historical Contrastive Category Discrimination (HCCD)
20
HCL: Semantic Segmentation Results
21
HCL: Semantic Segmentation Results
22
HCL: Semantic Segmentation Results
23
HCL: Semantic Segmentation Results
24
HCL: Semantic Segmentation Results
25
HCL: Image Classification Results
26
Thank you! Any questions?