CSE 252C: Selected Topics in Vision and Learning

Project Reports

June 1, 2011

Greg Long: Analyzing the Performance of the Bundler Algorithm

Abstract: The Bundler algorithm is a structure from motion algorithm rithm that produces point clouds from lists of unordered images. While Agarwal et al have demonstrated that the algorithm can produce accurate point clouds efficiently for very large datasets, they also list several ”research challenges,” including scale, variability of images, and accuracy. In this paper, we analyze the performance of the algorithm in these areas, using a variety of metrics.

Hani Altwaijry: Facial Attractiveness Classification

Abstract: The evaluation of attractiveness followed a scoring methodology rather than the usual classification one has not been attempted. This is desired for the sole reason that any person would be able to further decide between the attractiveness of two faces even if both are deemed attractive. In modelling this scenario, a continous scoring function is more appropriate. In this project, we implemented this scoring using Support Vector Machines to perform Regression on SIFT and Eigenfaces features extracted from the given data. The best results show a mean squared error of 0.082, which is considered promising in our case. With further refinement, such a tool will turn to be very useful in the hands of Internet dating services.

Andrew Ziegler: A Portable Structured Light System for the Rectification of Printed Text

Abstract: The goal of this project is to develop a portable structured light scanner which can capture the 3D structure of arbitrarily warped printed text and then rectify it. The resulting rectified text can then be sent to a standard OCR engine for recognition and layout analysis. The rectified image of the text can also be used as a reference image for non-rigid surface detection. Pairing non-rigid surface detection with the OCR results provides an infrastructure for augmented reality applications. The system is wearable and consists of a pico projector, a webcam, and a portable computer. This project is part of an on going effort to implement an assistive reading device for the visually impaired.

Oscar Beijbom: Active Learning and Confidence Prediction For Coral Reef Recognition

Abstract: Abstract We describe a series of experiments relating to patch size selection in a challenging dataset of coral reef images and show that we can achieve significant performance increase over the naive strategy of selecting a fixed patch size. Many computer vision methodologies use the bag of visual words representation of image objects. This is true, in particular, for the state of the art texture recognition frameworks. When applying these methods to recognition and segmentation of natural images, the issue of support region selection becomes crucial. We perform a series of experiments and conclude that the best strategy is to feed information aggregated over multiple support regions into a standard supervised machine learning technique.

Karan Sikka:


Rohan Anil: Bird’s eye view: Fine grained category recognition on the BIRDS 200 dataset

Abstract: Abstract In this project, we benchmark the Naive Bayes-Nearest Neighbor (NB-NN) algorithm on the Birds 200 dataset[6] using C-SIFT feature descriptor. Primary motivation is to provide a comprehensive set of baselines and experiments for further research on the dataset. In addition, we built a bird detector cascade for real time detection of birds in images and videos which is compatible with OpenCV. Finally, We explore on how to include side-information for which we use a log-linear model trained as a post processing step. Our preliminary results using NB-NN with C-SIFT descriptors on a subset of the dataset gives an accuracy of 25%.

Vicente Malave and Walter Talbott: A Shared Subspace for Multiple Metric Learning

Abstract: Many machine learning and computer vision problems (clustering, classification) make use of a distance. Starting with [20], it has been shown that it is possible to learn a suitably parametrized distance metric. For this project, we propose a new way of learning multiple metrics for the same dataset. We propose a formulation which shares dimensions of a common low-rank space. This metric not only allows us to pool information across categorization tasks, but can be used to understand which common stimulus dimensions are used by the algorithm.

Mohammad Moghimi: Using Color for Object Recognition

Abstract: Object Recognition is an important task in many Computer Vision problems. The problem is usually defined as classifying new images given a set of training images. One of the most challenging problems is Category-level Object Recognition which has got attention in the past decade. In this project the work on Fine-grained Category Object Recognition is presented. Despite General Object Recognition, images in this problem have a lot in common. But they still have differences that we can use for the classification. In this research, we use color feature for classification in many different ways. We are going to try to use color features which has not given much attention by the computer vision community to see if we can use it to improve the classification of 200 bird species (CUB200)[7]. The goal is to improve the performance of existing papers [5] [4] and also find more general way to use color in other similar problems.

Most recently updated on Oct. 29, 2007 by Serge Belongie.