Shuai ZhengTNT group meeting
1/12/2011
Paper Tracking Robust view transformation model for
gait recognition
Context-aware fusion: A case study on fusion of gait and face for human identification in video, 2010, Pattern Recognition.
Comments:This paper introduce how to combine
multi biometrics in context-aware way.Great summary for the existing work.New trends in long distance biometrics.
Effective Unconstrained Face Recognition by Combining Multiple Descriptors and Learned Background Statistics.2010, PAMI.
Comments:How to write a experimental paper?
That’s a model.
Cost-sensitive Face Recognition, Zhi-Hua Zhou, PAMI, 2010.
Comments:Good motivation: False identification,
false rejection, false acceptance are three different criteria, how to consider the whole cases together? To reduce the expectation of whole cost?
Multiclass cost-sensitive KLR seems the point of the paper.
Shuai Zheng, Junge Zhang, Kaiqi Huang, Tieniu Tan, Ran He.
MotivationMotivationMotivation from related work
Introduction Experimental results Conclusions and Future work
Robust gait representation should be robust to appearance variation caused by the change in viewing angle, carrying or wearing condition.
Shared gait representation subspace should be assumed as low-rank.
Handmade Low-Rank Truncated Singular Decomposition (TSVD) seems achieved better than original SVD in recent papers on multi-view gait recognition.
Robust low-rank method achieved exciting performance in background modeling, face recognition.
Related Work
We present a Robust View Transformation model and Partial Least Square feature selection algorithm for multi-view gait recognition.
Optimized GEI =
GEI from different views
Low-rank appx A+ Sparse error E
GEI
See? What a impressive results of robust View Transformation model for gait representation!
A Bag? Remove it as noise.
A overcoat? Remove it as noise.
The proposed method achieves significant performance on the multi-view gait recognition dataset with additional variations caused by wearing or carrying condition change.
sequelHow about the improved low-rank method for other challenge gait recognition dataset?
How about that for visual surveillance system?
Can we achieve super gait recognition? Achieved 99% recognition rates at any viewing angle? How about combine the method with rectified method?
No question? no reward!~
