alireza rahimpour, ali taalimi , hairong qi icassp 2017web.eecs.utk.edu/~arahimpo/poster.pdf ·...
TRANSCRIPT
RESEARCH POSTER PRESENTATION DESIGN © 2015
www.PosterPresentations.com
Motivation 2- Finding a Compact Model
q Distributed surveillance systems have become popular in recentyears due to security concerns.
q Transmitting the high dimensional data in bandwidth-limiteddistributed systems is a major challenge.
q We address this issue by proposing a novel probabilisticalgorithm based on the divergence between the probabilitydistributions of the visual features in order to reduce theirdimensionality.
1- Distributed Surveillance System
q Assume there are c classes in the training set and there are N feature histograms: in each class:
q Each bin of histogram is divided by the number of visual words ineach cluster and yields the probability density functions (pdf):
q The objective of this stage of the proposed approach is to compare allthe pdfs in each class and select a few informative ones by solving thefollowing optimization problem:
q After constructing a compact representation of the feature histograms for allthe classes in the training set (i.e., D), it will be saved in the smart cameras’memory.
q In the on-line testing stage in each of the p cameras, a feature histogram hi isextracted and encoded using the compact model D:
3- Generating Low Dimensional Feature Codes
FEATURE ENCODING IN BAND-LIMITED DISTRIBUTED SURVEILLANCE SYSTEMSAlireza Rahimpour, Ali Taalimi, Hairong Qi
Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville, TN, USAICASSP 2017
q Referred to as the Divergence-based Feature Selection (DFS) method.q 𝜔𝑖𝑗:the probability of 𝒇𝑖 being a representative for𝒇𝑗.
4- Experiments: Pedestrian Recognition in Surveillance Video
5- Experiments: Building Recognition in BMW Dataset
7- References6- Conclusion
Camera Setup [1] PRID dataset [1] Recognition accuracy for different dimensions of the feature histograms Confusion matrix of pedestrian recognition-340DPRIDdatasetusing340-Dfeatures.q Images contain a variety of viewpoints, illumination conditions, backgrounds and camera characteristics.
q We consider 30 different frames for each pedestrian in each camera for the baseline recognition.q At compression rate of 2.94, the accuracy is slightly better than using the original features. The reason is that our feature selection scheme omits
those features which are closer to the features from other classes than the features in their own class.
q The Berkeley Multi-view Wireless (BMW) database [2] consists of multiple-view images of 20landmark buildings on the Berkeley campus.
q Table 1 demonstrates the classification accuracy of different methods based on Sparse PCA (SPCA)[2] and Structure from Motion (SfM) [3].
q The physical interpretation of the reduced space is preserved during the dimensionality reductionprocedure, which is critical in the recognition task.
q The proposed probabilistic feature encoding approach achieves high compression rate while maintaining a high recognition accuracy.
q The DFS approach is applicable to variety of distributed computer vision tasks (e.g., cross view action recognition, person re-identification).
The Baseline Multi-view Recognition System:
q 𝐖 ∈ ℝ*×* ∶ the probability matrix for all the 𝒇-and 𝒇.pairs. When𝒇- is a representative for 𝒇., the corresponding row in the 𝐖 matrixis non-zero.
q The feature histograms corresponding to indices of non-zero rows of𝐖 are selected as our representative features in each class.
q Toy Example: finding the compact model for 3-class case:
[1]- Martin Hirzer, Csaba Beleznai, Peter M. Roth, and Horst Bischof, “Person re-identification bydescriptive and discriminative classification,” in (SCIA), 2011.[2]- Nikhil Naikal, Allen Y Yang, and S Shankar Sastry, “Informative feature selection for objectrecognition via sparse pca,” in IEEE International Conference on Computer Vision (ICCV), 2011.[3]- Panu Turcot and D Lowe, “Better matching with fewer features: The selection of usefulfeatures in large database recognition problems,” in ICCV workshop (WS-LAVD), 2009, vol. 4.Contact: [email protected]