Structure Based Analysis and Recognition in Images and Videos
Abstract
Statement of Work:Develop methods for (a) exploiting redundancy in image and video for object discovery, (b) image completion andlinear structure detection. They will evaluate these algorithms as follows.(1) Synthetic, corrupted, low-rank matrices. They will first generate multiple low-rank matrices of different dimensions, then corrupt them by sparse error matrices, and finally apply their method to recover the original low-rank matrices. They will attempt the same recovery using the existing methods and then compare those results with theirs in terms of running time and recovery accuracy. Since the ground truth of original low-rank matrices is known, they can evaluate the recovery accuracy quantitatively, such as in terms of Mean-of-Absolute-Error (MAE) and Peak-Signal-Noise-Ratio (PSNR).(2) Surveillance videos from the background subtraction database. (3) Face images from Extended Yale B database, which has been widely used to evaluate the low-rank recovery methods. For both (2) and (3), they will apply their method to recover low-rank components (images/videos) and compare running time/visual similarity with the state-ofthe-art (in quality) method RPCA. (4) Multi-resolution FTIR dataset. They will investigate the generality of their method on FTIR dataset, which places extreme demands on storage for both tissue and explosive materials, in terms ofefficiency of representation. Objective:This proposal addresses two important classes of problems in computer vision by developing methods that exploit intrinsic redundancy in images and videos and low-dimensionality of visual representations. More specifically, the PIs will develop methods for (a) exploiting redundancy in image/video for object discovery, and (b) image completion andlinear structure detection.Approach:The PI, Narendra Ahuja, addresses two important classes of problems in computer vision. The first is to extend recent advances in discovery, modeling and recognition to include images where contrast is low and hence segmentation is weak and unstable. They will extend object descriptors to include additional features and reduce the complexity. They propose to use low level geometric and photometric, 2D and 3D structure for predicting missing sub-images. Specifically, they will develop solutions to the problems of image and video completion. Second, they will investigate the problem of simplifying image and video representations by reducing dimensionality. They will first develop tools that efficiently and effectively identify redundancy within images and videos. They then propose to develop approaches to discovering linear structures in an unsupervised setting, to be used as models that explain the data, and developlinear models for prediction and recognition.Overall Merit and ONR Mission/Relevance:This research addresses Information Dominance and Autonomy focus areas. This work is expected to developcomponents for robust objection recognition and scene understanding from images and video.This research is expected to develop novel methods for dealing with occlusions in images for object recognition and scene understanding; develop improved objet descriptors for recognition; and a novel method for detection of explosives based on their infrared spectrum.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 08, 2016
- Source ID
- N000141612314
Entities
People
- Narendra Ahuja
Organizations
- Office of Naval Research
- United States Navy
- University of Illinois Urbana–Champaign