Cvss fall2012

Schedule Fall 2012

Date	Speaker	Title
September 6	Angjoo Kanazawa	Face Alignment by Explicit Shape Regression
September 13	Sameh Khamis	Combining Per-Frame and Per-Track Cues for Multi-Person Action Recognition
September 20	Douglas Summerstay	Artificial Intelligence and Artificial Creativity Before 1900
September 27	Mohammad Rastegari	Attribute Discovery via Predictable and Discriminative Binary Codes
October 4	Xavier Gibert Serra	Anomaly Detection on Railway Components using Sparse Representations
October 11	Kotaro Hara	Using Google Street View to Identify Street-level Accessibility Problems
October 18	Ashish Shrivastava	Dictionary learning methods for computer vision
October 25	Yi-Chen Chen	Dictionary-based Face Recognition from Video
November 1	Mohammad Rastegari	Instance Level Multiple Instance Learning Using Similarity Preserving Quasi Cliques
November 8	(Midterms, no meeting)
November 15	(CVPR deadline, no meeting)
November 22	(Thanksgiving, no meeting)
November 29	Fatemeh Mirrashed	Knowledge Adaptation in Visual Domains
December 6	Ejaz Ahmed	Linear Dimensionality Reduction Methods for Object Detection
December 13	Arijit Biswas	Clustering Images with Algorithms and Humans

Talk Abstracts Fall 2012

Face Alignment by Explicit Shape Regression

Speaker: Angjoo Kanazawa -- Date: September 6, 2012

In this talk, we will go over CVPR 2012 paper "Face Alignment by Explicit Shape Regression". I will review the paper and discuss its key concepts: cascaded regression, random ferns, shape indexed image features, and correlation based feature selection. Then I will discuss our hypothesis on why this seemingly simple method works so well and how we can apply their method to similar problem domains such as dog and bird parts localization and their challenges.

Abstract from the paper: We present a very efficient, highly accurate, “Explicit Shape Regression” approach for face alignment. Unlike previous regression-based approaches, we directly learn a vectorial regression function to infer the whole facial shape (a set of facial landmarks) from the image and explicitly minimize the alignment errors over the training data. The inherent shape constraint is naturally encoded into the regressor in a cascaded learning framework and applied from coarse to fine during the test, without using a fixed parametric shape model as in most previous methods. To make the regression more effective and efficient, we design a two-level boosted regression, shape-indexed features and a correlation-based feature selection method. This combination enables us to learn accurate models from large training data in a short time (20 minutes for 2,000 training images), and run regression extremely fast in test (15 ms for a 87 landmarks shape). Experiments on challenging data show that our approach significantly outperforms the state-of-the-art in terms of both accuracy and efficiency.

Combining Per-Frame and Per-Track Cues for Multi-Person Action Recognition

Speaker: Sameh Khamis -- Date: September 13, 2012

We propose a model to combine per-frame and per-track cues for action recognition. With multiple targets in a scene, our model simultaneously captures the natural harmony of an individual's action in a scene and the flow of actions of an individual in a video sequence, inferring valid tracks in the process. Our motivation is based on the unlikely discordance of an action in a structured scene, both at the track level (e.g., a person jogging then dancing) and the frame level (e.g., a person jogging in a dance studio). While we can utilize sampling approaches for inference in our model, we instead devise a global inference algorithm by decomposing the problem and solving the subproblems exactly and efficiently, recovering a globally optimal joint solution in several cases. Finally, we improve on the state-of-the-art action recognition results for two publicly available datasets.

Artificial Intelligence and Artificial Creativity Before 1900

Speaker: Doug Summers-Stay -- Date: September 20, 2012

I will talk about various inventions such as the Eureka, which generated Latin poetry in hexameter while playing "God Save the Queen"; the Homeoscope, a mechanical search engine invented by a Russian police clerk in 1832; the Componium, an orchestra-in-a-box which composed random variations on a melody; and others along the same lines. I'll also talk about how we could go beyond these techniques to build something really creative. This is a presentation of material I found when I was doing research for the book I published in January, Machinamenta.

Attribute Discovery via Predictable and Discriminative Binary Codes

Speaker: Mohammad Rastegari -- Date: September 27, 2012

We present images with binary codes in a way that balances discrimination and learnability of the codes. In our method, each image claims its own code in a way that maintains discrimination while being predictable from visual data. Category memberships are usually good proxies for visual similarity but should not be enforced as a hard constraint. Our method learns codes that maximize separability of categories unless there is strong visual evidence against it. Simple linear SVMs can achieve state-of-the-art results with our short codes. In fact, our method produces state-of-the-art results on Caltech256 with only 128- dimensional bit vectors and outperforms state of the art by using longer codes. We also evaluate our method on ImageNet and show that our method outperforms state-of-the-art binary code methods on this large scale dataset. Lastly, our codes can discover a discriminative set of attributes.

Anomaly Detection on Railway Components using Sparse Representations

Speaker: Xavier Gibert-Serra -- Date: October 4, 2012

High-speed rail (HSR) requires high levels of reliability of the track infrastructure. Automated visual inspection is useful for finding many anomalies such as cracks or chips on joint bars and concrete ties, but existing vision-based inspection systems often produce high number of false detections, and are very sensitive to external factors such as changes in environmental conditions. For example, state-of-the-art algorithms used by the railroad industry nominally perform at a detection rate of 85% with a false alarm rate of 3% and performance drops very quickly as image quality degrades. On the tie inspection problem, this false alarm rate would correspond to 2.6 detections per second at 125 MPH, which cannot be handled by an operator. These false detections have many causes, including variations in anomaly appearance, texture, partial occlusion, and noise, which existing algorithms cannot handle very well. To overcome these limitations, it is necessary to reformulate this joint detection and segmentation problem as a Blind Source Separation problem, and use a generative model that is robust to noise and is capable of handling missing data.

In signal and image processing, Sparse Representations (SR) is an efficient way of describing a signal as a linear combination of a small number of atoms (elementary signals) from a dictionary. In natural images, sparsity arises from the statistical dependencies of pixel values across the image. Therefore, statistical methods such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Independent Component Analysis (ICA) have been used for dimensionality reduction in several computer vision problems. Recent advances in SR theory have enabled methods that learn optimal dictionaries directly from training data. For example, K-SVD is a very well known algorithm for automatically designing over-complete dictionaries for sparse representation.

In this detection problem, the anomalies have very well defined structure and therefore, they can be represented sparsely in some subspace. In addition, the image background has very structured texture, so it is sparse with respect to a different frame. Theoretical results in mathematical geometric separation show that it is possible to separate these two image components (regular texture from contours) by minimizing the L1 norm the coefficients in geometrically complementary frames. More recently, it has been shown that this problem can be solved efficiently using thresholding and total variation regularization. Our experiments show that the sparse coefficients extracted from the contour component can be converted into feature vectors that can be used to cluster and detect these anomalies.

Using Google Street View to Identify Street-level Accessibility Problems

Speaker: Kotaro Hara -- Date: October 11, 2012

Poorly maintained sidewalks, missing curb ramps, and other obstacles pose considerable accessibility challenges; however, there are currently few, if any, mechanisms to determine accessible areas of a city a priori. In the first half of the presentation, I will talk about our investigation of the feasibility of using untrained crowd workers from Amazon Mechanical Turk (turkers) to find, label, and assess sidewalk accessibility problems in Google Street View imagery. Our work effectively demonstrates a promising new, highly scalable method for acquiring knowledge about sidewalk accessibility. In the latter half, I will discuss the future works as well as open research questions related in the field of computer vision.

Dictionary learning methods for computer vision

Speaker: Ashish Shrivastava -- Date: October 18, 2012

Sparse and redundant signal representations have recently gained much interest in image understanding. This is partly due to the fact that signals or images of interest are often sparse in some dictionary. These dictionaries can be either analytic or they can be learned directly from the data. In fact, it has been observed that learning a dictionary directly from data often leads to improved results in many practical applications such as classification and restoration. In this talk I will give a general overview of dictionary learning methods and talk in detail about my recent work on semi-supervised dictionary learning and non-linear supervised dictionary learning methods.

Dictionary-based Face Recognition from Video

Speaker: Yi-Chen Chen -- Date: October 25, 2012

The main challenge in recognizing faces in video is effectively exploiting the multiple frames of a face and the accompanying dynamic signature. One prominent method is based on extracting joint appearance and behavioral features. A second method models a person by temporal correlations of features in a video. Our approach introduces the concept of video-dictionaries for face recognition, which generalizes the work in sparse representation and dictionaries for faces in still images. Video-dictionaries are designed to implicitly encode temporal, pose, and illumination information. We demonstrate our method on the Face and Ocular Challenge Series (FOCS), which consists of unconstrained video sequences. We show that our method is efficient and performs significantly better than many competitive video-based face recognition algorithms.

Instance Level Multiple Instance Learning Using Similarity Preserving Quasi Cliques

Speaker: Mohammad Rastegari -- Date: November 1, 2012

In this work we introduce an instance-level approach to multiple instance learning. Our bottom-up approach learns a discriminative notion of similarity between instances in positive bags and use it to form a discriminative similarity graph. We then introduce the notion of similarity preserving quasi-cliques that aims at discovering large quasi-cliques with high scores of within-clique similarities. We argue that such large cliques provide clue to infer the underlying structure between positive instances. We use a ranking function that takes into account pairwise similarities coupled with prospectiveness of edges to score all positive instances. We show that these scores yield to positive instance discovery. Our experimental evaluations show that our method outperforms state-of-the-art MIL methods both at the bag-level and instance-level predictions in standard benchmarks and image and text datasets.

Knowledge Adaptation in Visual Domains

Speaker: Fatemeh Mirrashed -- Date: November 29, 2012

The new machine learning techniques of transfer learning and domain adaptation have recently captured special attention in the computer vision community. In this talk we will take a look at some of the methods that have been recently adopted or developed for adaptation of learning in the visual domains. We will also try to have an open discussion over some of more ideological questions such as better generalization versus adaptation. With abundance of massive volumes of visual training data should we keep at designing algorithms that could model all the possible variations in the visual world or should we regard adaptation as an integral part of learning in the visual domains?

Linear Dimensionality Reduction Methods for Object Detection

Speaker: Ejaz Ahmed -- Date: December 6, 2012

I'll talk about linear methods for dimensionality reduction, emphasizing on PLS and CDF (composite discriminant factors). Focus on linear methods is motivated by the task of object detection which can be benefited from linear projections in various ways. I'll go through some of these issues as well.

We propose a linear dimensionality reduction method, Composite Discriminant Factor (CDF) analysis, which searches for a discriminative but compact feature subspace that can be used as input to classifiers that suffer from problems such as multi-collinearity or the curse of dimensionality. The subspace selected by CDF maximizes the performance of the entire classification pipeline, and is chosen from a set of candidate subspaces that are each discriminative by various local measures, such as covariance between input features and output labels or the margin between positive and negative samples. Our method is based on Partial Least Squares (PLS) analysis, and can be viewed as a generalization of the PLS1 algorithm, designed to increase discrimination in classification tasks. While our experiments focus on improvements to object detection (in particular, pedestrians and vehicles), a task that often involves high dimensional features and benefits from fast linear approaches, we also demonstrate our approach on machine learning datasets from the UCI Machine Learning repository. Experimental results show that the proposed approach improves significantly over SVM in terms of accuracy and also over PLS in terms of compactness and efficiency, while maintaining or slightly improving accuracy.

Clustering Images with Algorithms and Humans

Speaker: Arijit Biswas -- Date: December 13, 2012

In this talk, we address the problem of image clustering. Although there are many excellent clustering algorithms, effective clustering remains very challenging for datasets that contain many classes. We propose several approaches to cluster images accurately. First, we use pairwise constraints from humans to cluster images. An algorithm provides us with pairwise image similarities. We then actively obtain selected, more accurate pairwise similarities from humans. A novel method is developed to choose the most useful pairs to show a person, obtaining constraints that improve clustering. Second, we propose a new algorithm to cluster a subset of the images only (we call this subclustering), which will produce a few examples from each class. Subclustering will produce smaller but purer clusters. Finally, we make use of human input in an active subclustering algorithm to further improve results.

Web Accessibility