About AMP Lab        Projects        Downloads        Publications        People        Links

Group Member

Devi Parikh
Graduate Student

Personal Homepage:   http://ttic.uchicago.edu/~dparikh/

Email: dparikh@cmu.edu


Office: Hamerschlag Hall D-level

             Cubicle A3

Fax: 412-268-3890

Mailing Address:
ECE Department,

1111 Hamerschlag Hall

Carnegie Mellon University,
5000 Forbes Avenue
Pittsburgh, PA 15213-3890

[Research Interests]        [Projects]      [Publications

Research Interests

  Top of this page




Low-level Contextual Patch Saliency

Context has been extensively explored for higher level tasks such as recognition. In this work we explore the use of context for a low-level task of determining patch saliency.

Existing saliency measures include interest-points based and discriminative criteria. We define saliency of a patch based on consistency of the patch in context of the rest of the image. We propose occurrence as well as location based contextual saliency measures. We also explore three different strategies to sample patches from a saliency map.

We evaluate our proposed contextual saliency measure by comparing it to existing saliency measures for the tasks of object and scene recognition.

 Please refer to Low-level Contextual Patch Saliency for more details.


From Appearance to Context Based Recognition

One can imagine that contextual information is necessary for accurate recognition in low resolution images. However, the same is not the case for high resolution images.

To verify this, in addition to automatic machine recognition, we perform recognition tests in human studies as well.

In this paper, we explore the role of context for dense scene labeling in small images. We promote the scenario with impoverished appearance information as an appropriate venue for studying the role of context in recognition.

 Please refer to From Appearance to Context Based Recognition for more details.



Hierarchical Semantics of Objects (hSO)


We introduce hSOs: Hierarchical Semantics of Objects. The hSO captures the interactions between the objects that tend to co-occur in the scene, and hence are potentially semantically related. The proposed approach is entirely unsupervised and can detect the parts of the images that belong to the foreground objects, cluster these parts to represent objects, and provide an understanding of the scene by hierarchically clustering these objects in a semantically meaningful way - all from a collection of unlabeled images of a particular scene category. In addition to providing the semantic layout of the scene, learnt hSOs can have several useful applications such as compact scene representation for scene category classification and providing context for enhanced object detection. 

Please refer to Hierarchical Semantics of Objects (hSO) for more details.




Pattern Recognition Techniques for Intrusion Detection Systems


We propose a pattern recognition based novel strategy for adaptive intrusion detection that can evolve with changing network environments. We exploit the ensemble of classifiers approach to combine information from multiple sources and tune the system towards minimizing the cost of the errors. We also propose a novel transform that can map the outputs of different classifiers to common grounds for more meaningful comparison/combination.


Please refer to Pattern Recognition Techniques for Intrusion Detection Systems for more details.






3D Reassembly

We propose a local feature-based approach to determine compatibility between parts for 3D reassembly. We use a spectral technique to compute the compatibility score between parts.


Please refer to 3D reassembly for more details.







Top of this page


Journal Papers:

Conference Papers: 

Top of this page

free hit counter

This website is maintained by Devi Parikh
Copyright © 2001 
Advanced Multimedia Processing Lab. All rights reserved.
Revised: January 26, 2009 .