Cvss fall2013
Schedule Fall 2013[edit]
All talks take place on Thursdays at 4:30pm in AVW 3450.
| Date | Speaker | Title |
|---|---|---|
| September 19 | Mohammad Rastegari | Fast Image Prior |
| September 26 | (no meeting) | |
| October 3 | (MSR talk, no meeting) | |
| October 10 | Yezhou Yang | A Context-free Manipulation Action Grammar and Manipulation Action Consequences Detection |
| October 17 | Garrett Warnell | Ray Saliency: Bottom-Up Saliency for a Rotating and Zooming Camera |
| October 24 | Abhishek Sharma | A Sentence is Worth a Thousand Pixels |
| October 31 | (CVPR deadline, no meeting) | |
| November 7 | Jingjing Zheng | Cross-View Action Recognition Via a Transferable Dictionary Pair |
| November 14 | Sumit Shekhar | Joint Sparse Representation for Multimodal Biometric Recognition |
| November 21 | Jonghyun Choi | Renaissance of Convolutional Neural Network - what, why and so? |
| November 28 | (Thanksgiving, no meeting) | |
| December 5 | Arijit Biswas | Distance Learning Using the Triangle Inequality for Semi-supervised Clustering |
Talk Abstracts Fall 2013[edit]
Fast Image Prior[edit]
Speaker: Mohammad Rastegari -- Date: September 19, 2013
In this project we introduce a new method for learning image prior that can be used for many applications in image reconstruction. We learn a generative model on natural image patches. Our generative model is similar to one in Gausian Mixture Model (GMM). The key idea of our approach is to force each component of our generative model to share the same set of basis vectors. This leads to a much faster inference at test time. We used image denoising as our test bed for this image prior learning. Our experimental results shows that we reached about 30x speed up over state-of-the-art method while getting slightly improvement in denoising accuracy.
A Context-free Manipulation Action Grammar and Manipulation Action Consequences Detection[edit]
Speaker: Yezhou Yang -- Date October 10, 2013
Humanoid robots will need to learn the actions that humans perform. They will need to recognize these actions when they see them and they will need to perform these actions themselves. In this presentation I will introduce a manipulation grammar to perform this learning task. Context-free grammars in linguistics provide a simple and precise mechanism for describing the methods by which phrases in some natural language are built from smaller blocks. Also, the basic recursive structure of natural languages is described exactly. Similarly, for manipulation actions, every complex activity is built from smaller blocks involving hands and their movements, as well as objects, tools and the monitoring of their state. Thus, interpreting a seen action is like understanding language, and executing an action from knowledge in memory is like producing language. Associated with the grammar, a parsing algorithm is proposed, which can be used bottom-up to interpret videos by dynamically creating a semantic tree structure, and top-down to create the motor commands for a robot to execute manipulation actions. Experiments on both tasks, i.e. a robot observing people performing manipulation actions, and a robot executing manipulation actions on a simulation platform, validate the proposed formalism.
Ray Saliency: Bottom-Up Saliency for a Rotating and Zooming Camera[edit]
Speaker: Garrett Warnell -- Date: October 17, 2013
We extend the classical notion of visual saliency to multi-image data collected using a stationary pan-tilt-zoom (PTZ) camera. We show why existing saliency methods are not effective for this type of data, and propose ray saliency: a modified notion of visual saliency that utilizes knowledge of the imaging process in order to appropriately incorporate the context provided by multiple images. We present a practical, mosaic-free method by which to quantify and calculate ray saliency, and demonstrate its usefulness on PTZ imagery.
A Sentence is Worth a Thousand Pixels[edit]
Speaker: Abhishek Sharma -- Date: October 24, 2013
We are interested in holistic scene understanding where images are accompanied with text in the form of complex sentential descriptions. We propose a holistic conditional random field model for semantic parsing which reasons jointly about which objects are present in the scene, their spatial extent as well as semantic segmentation, and employs text as well as image information as input. We automatically parse the sentences and extract objects and their relationships, and incorporate them into the model, both via potentials as well as by re-ranking candidate detections. We demonstrate the effectiveness of our approach in the challenging UIUC sentences dataset and show segmentation improvements of 12.5% over the visual only model and detection improvements of 5% AP over deformable part-based models.
Cross-View Action Recognition Via a Transferable Dictionary Pair[edit]
Speaker: Jingjing Zheng -- Date: November 7, 2013
Discriminative appearance features are effective for recognizing actions in a fixed view, but generalize poorly to changes in viewpoint. We present a method for view-invariant action recognition based on sparse representations using a transferable dictionary pair. A transferable dictionary pair consists of two dictionaries that correspond to the source and target views respectively. The two dictionaries are learned simultaneously from pairs of videos taken at different views and aim to encourage each video in the pair to have the same sparse representation. Thus, the transferable dictionary pair links features between the two views that are useful for action recognition. Both unsupervised and supervised algorithms are presented for learning transferable dictionary pairs. Using the sparse representation as features, a classifier built in the source view can be directly transferred to the target view. We extend our approach to transferring an action model learned from multiple source views to one target view. We demonstrate the effectiveness of our approach on the multi-view IXMAS data set. Our results compare favorably to the the state of the art.
Joint Sparse Representation for Multimodal Biometric Recognition[edit]
Speaker: Sumit Shekhar -- Date: November 14, 2013
In this talk, I will present the work on feature-level fusion method for multimodal biometric recognition. Traditional methods for combining outputs from different modalities are based on score-level or decision-level fusion. Feature-level fusion can be more discriminative, but has hardly been explored due to challenges of different feature outputs and high feature dimensions. Here, I will present a framework using joint sparsity to combine information, and show its application to multimodal biometric recognition, face recognition and vidoe-based recognition.
Renaissance of Convolutional Neural Network - what, why and so?[edit]
Speaker: Jonghyun Choi -- Date: November 21, 2013
The convolutional neural network based deep networks recently improve image classification accuracy significantly over the state-of-the-art vision approaches. I will go through what the successful deep convolutional neural net looks like, why it is again popular now and on-going deep net research in other research groups. I will mostly go through the successful instance of deep convolutional neural net tuned by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, published in NIPS 2012.
Distance Learning Using the Triangle Inequality for Semi-supervised Clustering[edit]
Speaker: Arijit Biswas -- Date: December 5, 2013
Success of semi-supervised clustering algorithms depends on how effectively supervision can be propagated to the unsupervised data. We propose a method for modifying all pairwise image distances when must-link or can't-link pairwise constraints are provided for only a few image pairs. These distances are used for clustering images. First, we formulate a brute-force Quadratic Programming (QP) method that modifies the distances such that the total change in distances is minimized but the final distances obey the triangle inequality. Then we propose a much faster version of the QP that can be applied to large datasets by enforcing only a selected subset of the inequalities. We prove that this still ensures that key qualitative properties of the distances are correctly computed. We run experiments on face, leaf and video image clustering and show that our proposed approach outperforms state-of-the-art methods for constrained clustering.
