Pedestrian Attribute Recognition (PAR) has received extensive attention during the past few years. With the advances of deep convolutional neural networks (CNNs), the performance of PAR has been significantly improved. Existing methods tend to acquire attribute-specific features by designing various complex network structures with additional modules. Such additional modules, however, dramatically increase the number of network parameters. Meanwhile, the problems of class imbalance and hard attribute retrieving remain underestimated in PAR. In this paper, we explore the optimization mechanism of the training processing to account for these problems and propose a new loss function called Multi-label Contrastive Focal Loss (MCFL). This proposed MCFL emphasizes the hard and minority attributes by using a separated re-weighting mechanism for different positive and negative classes to alleviate the impact of the imbalance. MCFL is also able to enlarge the gaps between the intra-class of multi-label attributes, to force CNNs to extract more subtle discriminative features. We evaluate the proposed MCFL on three large public pedestrian datasets, including RAP, PA-100K, and PETA. The experimental results indicate that the proposed MCFL with the ResNet-50 backbone is able to outperform other state-of-the-art approaches in term of mean accuracy.

Many recent machine learning models rely on fine-grained dynamic control flow for training and inference. In particular, models based on recurrent neural networks and on reinforcement learning depend on recurrence relations, data-dependent conditional execution, and other features that call for dynamic control flow. These applications benefit from the ability to make rapid control-flow decisions across a set of computing devices in a distributed system. For performance, scalability, and expressiveness, a machine learning system must support dynamic control flow in distributed and heterogeneous environments.

Analysis of degenerate tensors is a fundamental step in finding the topological structures and separatrices in tensor fields. Previous work in this area have been limited to analyzing symmetric second order tensor fields. In this paper, we extend the topological analysis to 2D general (asymmetric) second order tensor fields. We show that it is not sufficient to define degeneracies based on eigenvalues alone, but one must also include the eigenvectors in the analysis. We also study the behavior of these eigenvectors as they cross from one topological region into another.

Tensor topology is useful in providing a simplified and yet detailed representation of a tensor field. Recently the field of 3D tensor topology is advanced by the discovery that degenerate tensors usually form lines in their most basic configurations. These lines form the backbone for further topological analysis. A number of ways for extracting and tracing the degenerate tensor lines have also been proposed. In this paper, we complete the previous work by studying the behavior and extracting the separating surfaces emanating from these degenerate lines. First, we show that analysis of eigenvectors around a 3D degenerate tensor can be reduced to 2D. That is, in most instances, the 3D separating surfaces are just the trajectory of the individual 2D separatrices which includes trisectors and wedges. But the proof is by no means trivial since it is closely related to perturbation theory around a pair of singular slate. Such analysis naturally breaks down at the tangential points where the degenerate lines pass through the plane spanned by the eigenvectors associated with the repeated eigenvalues. Second, we show that the separatrices along a degenerate line may switch types (e.g. trisectors to wedges) exactly at the points where the eigenplane is tangential to the degenerate curve. This property leads to interesting and yet complicated configuration of surfaces around such transition points. Finally, we apply the technique to several common data sets to verify its correctness.

This paper addresses several issues related to topological analysis of 3D second order symmetric tensor fields. First, we show that the degenerate features in such data sets form stable topological lines rather than points, as previously thought. Second, the paper presents two different methods for extracting these features by identifying the individual points on these lines and connecting them. Third, this paper proposes an analytical form of obtaining tangents at the degenerate points along these topological lines. The tangents are derived from a Hessian factorization technique on the tensor discriminant and leads to a fast and stable solution. Together, these three advances allow us to extract the backbone topological lines that form the basis for topological analysis of tensor fields.

Visualization of 3D tensor fields continues to be a major challenge in terms of providing intuitive and uncluttered images that allow the users to better understand their data. The primary focus of this paper is on finding a formulation that lends itself to a stable numerical algorithm for extracting stable and persistent topological features from 2nd order real symmetric 3D tensors. While features in 2D tensors can be identified as either wedge or trisector points, in 3D, the corresponding stable features are lines, not just points. These topological feature lines provide a compact representation of the 3D tensor field and are essential in helping scientists and engineers understand their complex nature. Existing techniques work by finding degenerate points and are not numerically stable, and worse, produce both false positive and false negative feature points. This work seeks to address this problem with a robust algorithm that can extract these features in a numerically stable, accurate, and complete manner.

We introduce a new method for visualizing symmetric tensor fields. The technique produces images and animations reminiscent of line integral convolution (LIC). The technique is also slightly related to hyperstreamlines in that it is used to visualize tensor fields. However, the similarity ends there. HyperLIC uses a multi-pass approach to show the anisotropic properties in a 2D or 3D tensor field. We demonstrate this technique using data sets from computational fluid dynamics as well as diffusion-tensor MRI.

Visualizing second-order 3D tensor fields continue to be a challenging task. Although there are several algorithms that have been presented, no single algorithm by itself is sufficient for the analysis because of the complex nature of tensor fields. In this paper, we present two new methods, based on volume deformation, to show the effects of the tensor field upon its underlying media. We focus on providing a continuous representation of the nature of the tensor fields. Each of these visualization algorithms is good at displaying some particular properties of the tensor field.

Generalized partially-mixed-polarity Reed-Muller (GPMPRM) expansion, a canonical subfamily of exclusive sum of products (ESOP), is presented. An efficient algorithm in two-dimensional data flow is proposed for computation of the GPMPRM forms. MCNC benchmark experimental results show that the minimal GPMPRM forms of these functions, on the average, have similar number of terms to their sum of products (SOP) counterparts while there are many functions for which the GPMPRM circuits are much smaller.