Analysis-Synthesis Dictionary Learning for Universality-Particularity Representation Based Classification
Authors: Meng Yang, Weiyang Liu, Weixin Luo, Linlin Shen
AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we evaluate the performance of ASDL-UP on various classification task. Action classification, face recognition, gender classification are performed by using ASDL-UP and the competing methods in the following Sections. |
| Researcher Affiliation | Academia | College of Computer Science & Software Engineering, Shenzhen University, China School of Electronic & Computer Engineering, Peking University, China |
| Pseudocode | Yes | Algorithm 1 Training Procedure of ASDL-UP |
| Open Source Code | No | More experimental and time complexity analysis are presented in the supplementary material1. 1available at www.yangmeng.org.cn or www.wyliu.com |
| Open Datasets | Yes | The benchmark action dataset, UCF sports action dataset (M. Rodriguez 2008), is used to conduct the action classification experiment. ... UCF50 action dataset by using fivefold group-wise cross validation, and compared it with the recent DL methods and the other state-of the-art methods, including (S. Sadanand 2012) and (H.R. Wang 2012). ... Two face databases, such as the aligned labeled face in the wild (LFWa)(L. Wolf 2009) and AR (A. Martinez 1998), are used to evaluate the performance of the proposed ASDL-UP. |
| Dataset Splits | Yes | As the experiment setting in (Z.L. Jiang 2013), we evaluated the ASDL-UP via five-fold cross validation. Here the dimension of the action bank feature is reduced to 100 via PCA... Following the experiment settings in (S. Sadanand 2012), we then evaluated ASDL-UP on the large-scale UCF50 action dataset by using fivefold group-wise cross validation... 20 images of each subject are used for training and the remaining 6 images are used for testing. ... For each person the first 10 samples are used for training data with the remaining samples for testing. ... We randomly select 585 male samples and 622 female samples as the training set, while the remaining 11495 face images used as the test data. |
| Hardware Specification | No | The average running times on two datasets using Matlab 2011a are listed in Table 6. |
| Software Dependencies | No | The average running times on two datasets using Matlab 2011a are listed in Table 6. |
| Experiment Setup | Yes | There are three parameters,λ,γ, and η, in the model of ASDL-UP (i.e.,Eq.(4)). λ is a parameter of discriminative class-particular analysis coding coefficient term, γ is a parameter of discriminative universal analysis coefficient term, and η is a parameter of hard-thresholding function. In our experiments, we fix λ = 1e 3 in all experiments and set γ = 0.1, and η = 1e 5 for face recognition, and γ = 0.5, and η = 1e 4 for all the other experiments. |