Training Deep neural networks (DNNs) on noisy labeled datasets is a challenging problem, because learning on mislabeled examples deteriorates the performance of the network. As the ground truth availability is limited with real-world noisy datasets, previous papers created synthetic noisy datasets by randomly modifying the labels of training examples of clean datasets. However, no final conclusions can be derived by just using this random noise, since it excludes feature-dependent noise. Thus, it is imperative to generate feature-dependent noisy datasets that additionally provide ground truth. Therefore, we propose an intuitive approach to creating feature-dependent noisy datasets by utilizing the training predictions of DNNs on clean datasets that also retain true label information. We refer to these datasets as "Pseudo Noisy datasets". We conduct several experiments to establish that Pseudo noisy datasets resemble feature-dependent noisy datasets across different conditions. We further randomly generate synthetic noisy datasets with the same noise distribution as that of Pseudo noise (referred as "Randomized Noise") to empirically show that i) learning is easier with feature-dependent label noise compared to random noise, ii) irrespective of noise distribution, Pseudo noisy datasets mimic feature-dependent label noise and iii) current training methods are not generalizable to feature-dependent label noise. Therefore, we believe that Pseudo noisy datasets will be quite helpful to study and develop robust training methods.

The MPEG committee has completed development of a new audio coding standard called “MPEG-4 advanced audio coding-enhanced low delay” (AAC-ELD). AAC-ELD uses low delay spectral band replication (LD-SBR) technology together with a low delay time domain alias cancellation (LD TDAC) filterbank in the encoder to achieve both high coding efficiency and low algorithmic delay. In this paper, we present fast algorithms for implementing LD-TDAC filterbanks in AAC-ELD. Two types of fast algorithms are presented. In the first, we map LD-TDAC analysis and synthesis filterbanks to modified discrete cosine transform (MDCT) and inverse modified discrete cosine transform (IMDCT), respectively. Since MDCT/IMDCT are already extensively used in AAC and they have many fast algorithms, this mapping not only provides a fast implementation but also allows a common implementation of the filterbanks in AAC Low Complexity (AAC-LC), AAC Low Delay (AAC-LD) and AAC-ELD codecs. In the second algorithm, we provide a mapping to discrete Cosine transform of type II. The mapping to DCT-II allows the merger of the matrix operations with the windowing stage that precedes or follows them. This further reduces the number of multiplications and leads to an algorithm with the lowest known arithmetic complexity. For filterbanks of lengths 1024 and 960, we also present a new fast factorization of 15-point DCT-II that requires only 14 irrational multiplications, 3 dyadic rational multiplications and 67 additions.

The MPEG committee has recently completed development of a new audio coding standard “MPEG-4 Advanced Audio Coding-Enhanced Low Delay” (AAC-ELD). AAC-ELD is targeted towards high-quality, full-duplex communication applications such as audio and video conferencing. AAC-ELD uses low delay spectral band replication (LD-SBR) technology together with a low delay AAC core encoder to achieve high coding efficiency and low algorithmic delays. In this paper, we present fast algorithms for computing LD-SBR filterbanks in AAC-ELD. The proposed algorithms map complex exponential modulation portion of the filterbanks to discrete cosine transforms of types IV and II. Our proposed mapping also allows to merge some multiplications with the windowing stage that precedes or succeeds the modulation step. This further reduces computational complexity. Our presentation includes detailed explanation and flow-graphs of the algorithms, complexity analysis, and comparisons with alternative implementations.

One of the key elements of the data integration process is conflation of spatial data from two different sources, which results in a single data set that incorporates data and attributes from both sources. Registration techniques needed for conflation can be rigid or non-rigid, depending on the underlying transformation. A key characteristic of rigid transformation is that all distances are preserved.

An efficient and robust framework is proposed for two-view multiple structure-and-motion segmentation of unknown number of rigid objects. The segmentation problem has three unknowns, namely the object memberships, the corresponding fundamental matrices, and the number of objects. To handle this otherwise recursive problem, hypotheses for fundamental matrices are generated through local sampling. Once the hypotheses are available, a combinatorial selection problem is formulated to optimize a model selection cost which takes into account the hypotheses likelihoods and the model complexity. An explicit model for outliers is also added for robust segmentation. The model selection cost is minimized through the branch-and-bound technique of combinatorial optimization. The proposed branch-and-bound approach efficiently searches the solution space and guarantees optimality over the current set of hypotheses. The efficiency and the guarantee of optimality of the method is due to its ability to reject solutions without explicitly evaluating them. The proposed approach was validated with synthetic data, and segmentation results are presented for real images.

Segmentation problems are one of the most important areas of research in computer vision. While segmentation problems are generally solved with clustering paradigms, they formulate the problem as recursive. Additionally, most approaches need the number of clusters to be known beforehand. This requirement is unreasonable for majority of the computer vision problems. This paper analyzes the model selection perspective which can overcome these limitations. Under this framework multiple hypotheses for cluster centers are generated using spatially coherent sampling. An optimal subset of these hypotheses is selected according to a model selection criterion. The selection can be carried out with a branch-and-bound procedure. The worst case complexity of any branch-and-bound algorithm is exponential. However, the average complexity of the algorithm is significantly lower. In this paper, we develop a framework for analysis of average complexity of the algorithm from the statistics of model selection costs.

A split-and-merge framework based on a maximum variance criterion is proposed for disparity clustering. The proposed algorithm transforms low-level stereo disparity information to mid-level planar surface information which can be used further to carry out high-level computer vision tasks such as shape classification. Unlike conventional clustering, the proposed algorithm assumes that the number of clusters is unknown. Instead, a maximum variance criterion is applied to extract planar surfaces from the disparity image. The split phase of the algorithm creates clusters based on spatial continuity and the merge phase combines these clusters such that variance per cluster does not exceeded an allowable value. For efficient maximum variance clustering, a greedy branch-and-bound procedure is introduced. Efficiency of the approach is verified through experiments.

An iterative segmentation-estimation framework for segmentation of planar surfaces in the disparity space is proposed. Disparity of a scene is modeled by approximating various surfaces in the scene to be planar. Uncertainty in the stereo disparity of these surfaces is modeled with a hidden Markov random field. Planar surface labels are the hidden variables of the model. Surface labels are estimated during the segmentation phase of the framework with help of underlying plane parameters and the spatial coherency prior. After segmentation, the planar surfaces are separated into spatially continuous regions. Each spatially continuous region is treated as a hypothesis for the underlying plane. Hypotheses for each plane are then combined to form a multihypothesis Gaussian mixture prior. The underlying planar surface parameters are estimated with maximum a posteriori estimation from the multihypothesis prior. The iterative process is continued until the labels converge or satisfactory segmentation is achieved. Experimental results are presented with synthetic as well as real-life scenes to demonstrate the success of the proposed method.

An iterative split-and-merge framework for the segmentation of planar surfaces in the disparity space is presented. Disparity of a scene is modeled by approximating various surfaces in the scene to be planar. In the split phase, the number of planar surfaces along with the underlying plane parameters is assumed to be known from the initialization or from the previous merge phase. Based on these parameters, planar surfaces in the disparity image are labeled to minimize the residuals between the actual disparity and the modeled disparity. The labeled planar surfaces are separated into spatially continuous regions which are treated as candidates for the merging that follows. The regions are merged together under a maximum variance constraint while maximizing the merged area. A multistage branch-and-bound algorithm is proposed to carry out this optimization efficiently. Each stage of the branch-and-bound algorithm separates a planar surface from the set of spatially continuous regions. The multistage merging estimates the number of planar surfaces and their labeling. The splitting and the multistage merging is repeated till convergence is reached or satisfactory results are achieved. Experimental results are presented for variety of stereo image data.

A new relationship is derived between blocking probability and the rearrangement probability density for a three stage switch. A novel analytical model for call packing is developed. The model confirms previous results from simulation and shows that packing/repacking reduces blocking and hence the required rearrangements, compared to the random routing method. The probability density of rearrangements is also derived.>

New adaptive differential pulse code modulation (ADPCM) coders with adaptive prediction are proposed and compared with existing nonadaptive DPCM coders, for processing composite National Television System Commission (NTSC) television signals. Comparisons are based on quantative criteria as well as subjective evaluation of the processed still frames. The performance of the proposed predictors is shown to be independent of well-designed quantizers and better than existing predictors in such critical regions of the pictures as edges and contours.