1 / 25

today

today. you will entertain me. Sender. Enhancing the Throughput of Video Streaming Using Automatic Colorization. Internet. Receiver. Automatic Colorization. Scene Recognition Improvement. GIST. Spatial Pyramid. Histogram on Geometry Segmentation Regions.

aislin
Download Presentation

today

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. today

  2. you will entertain me

  3. Sender Enhancing the Throughput of Video Streaming Using Automatic Colorization Internet Receiver Automatic Colorization

  4. Scene Recognition Improvement GIST Spatial Pyramid Histogram on Geometry Segmentation Regions J. Xiao, J. Hays, K. Ehinger, A. Oliva, and A. Torralba SUN Database: Large-scale Scene Recognition from Abbey to Zoo CVPR2010

  5. Content-Awareness Text Recognition Linear Algebra Lecture 21 Eigenvalues – Eigenvectors Det [A - λI] = 0 TRACE = λ1 + λ2 + … + λn 1. A piece of lecture video 2. Identify the text region

  6. Fine-grained RF localization and video matting for privacy-preserving “opt-in” video services and dynamic layers Our first attempt at a privacy system was a blunt instrument– press the big red button on your personal tag to kill the sensors. There are two big problems with this approach: it is strictly “opt-out,” and one person’s preferences affect everyone around them. My research group recently deployed an invasive sensor network to investigate issues such as privacy and narrative in the era of ubiquitous surveillance. The next step is to create an opt-in system, in which users who wish to be seen carry a special RF tag and everyone else is invisible (replaced by background). This is made possible by recent advances in RF localization and computer vision.

  7. Fine-grained RF localization and video matting for privacy-preserving “opt-in” video services and dynamic layers The first step is to implement the computer vision services necessary to make such a system work: Background estimation & video matting Motion estimation for image-tag correspondence 3-d from a single camera to deal with occlusions Correspondence across cameras with overlapping views

  8. vs. Generative Models for Texture SynthesisLawson Wong • Goals of generative-model based vision: (GMBV Workshop at CVPR 2004) • Formulate a model of image generation • Estimate posterior probability of model parameters given image • Applications: Object detection, recognition, tracking … • All of vision! We can solve everything! • But we haven’t solved everything yet • Perhaps we should look at a simpler case • Proposal: Textures • Much more structure than images • Possibly forms a basic unit for understanding images • Project goal: Model and synthesize textures

  9. A Study of Motion Layer Segmentation Algorithms • In this project I intend to experiment with motion layer segmentation algorithms. The purpose of motion segmentation is, given an image sequence, to decompose the sequence into layers of pixels, moving coherently through time as a result of an underlying process. Such algorithms typically operate on top of dense motion field estimation, and so it should be interesting to try and apply them using different optical flow algorithms. Another challenge lies on the actual methodology used to evaluate their results. To my knowledge, a comprehensive study that compares the results of motion segmentation algorithms to ground truth data has yet to be conducted (the only reference I found is [7]). • The project will probably involve implementing 1-2 such algorithms and investigating their performance on synthetic and natural video data. I would like to start with the fundamentals, and have already implemented the simple approach described in [1]. I would also like to experiment with [2]/[3]/[4] and if time will allow with newer approaches such as [5]/[6]. Some references: [1] Y. Weiss, Motion Segmentation using EM - a short tutorial [2] J. Wang, E. Adelson, Representing Moving Images with Layers, ToIP 94 [3] Y. Weiss, E. Adelson, Perceptually organized EM: A framework for motion segmentation that combines information about form and motion, ICCV 95 [4] J. Shi, J. Malik, Motion Segmentation and Tracking Using Normalized Cuts, ICCV 98 [5] Q. Ke, T. Kanade, A Robust Subspace Approach to Layer Extraction, IEEE motion 02, [6] M. Kumar, P. Torr, A. Zisserman, Learning Layered Motion Segmentations of Video, IJCV 07 [7] L. Zappella et al., Motion Segmentation: a Review, Proceeding of the 2008 conference on Artificial Intelligence Research and Development 08

  10. Roarke Horstmeyer Computer Vision Final Project: Does multispectral imaging improve object recognition and segmentation? Concept: There are many current approaches using RGB color values (sometimes in another space like YUV) to perform object segmentation and help with texture segmentation,. Also, RGB color content is combined with position info in algorithms like Mean Shift Segmentation. Hypothesis: Can using more than 3 overlapping color measurements improve segmentation algorithms? What is performance vs. # spectral bands? Approach: a. Modify k-means based algorithm, mean shift segmentation, and other algorithms to use additional spectral content. b. Test algorithms with 30 spectral channel shared database: http://www.cs.columbia.edu/CAVE/databases/multispectral/ c. Take own data with multiple spectral filters using a 5x5 camera array More colors Improved segmentation?

  11. SIFT On the GPU

  12. Automated Prediction of Consumer Response From Face and Gestures • Which actions and gestures are salient, which are not? • Can we determine from fewer tests whether the product will be a hit? Track Features Can we predict the outcome of consumer tests from facial actions and gestures? Identify Actions/ Gestures • METHODOLOGY: • Track features • Identify actions/gestures • Map to response – like/dislike • Validate against test set • Evaluate performance Predict Response

More Related