This paper presents a multi-output regression model for crowd counting in public scenes. Existing by methods either learn single global counting, or train large number of separate regressors localised density estimation. In contrast, our based approach is able to estimate people count spatially regions and more scalable without the need training proportional local regions. particular, proposed automatically learns functional mapping between interdependent low-level features multi-dimensional...

In this paper, we investigate the issue of knowledge distillation for training compact semantic segmentation networks by making use cumbersome networks. We start from straightforward scheme, pixel-wise distillation, which applies scheme originally introduced image classification and performs each pixel separately. further propose to distill structured into networks, is motivated fact that a prediction problem. study two such schemes: (i) pair-wise distills pairwise similarities, (ii)...

A number of computer vision problems such as human age estimation, crowd density estimation and body/face pose (view angle) can be formulated a regression problem by learning mapping function between high dimensional vector-formed feature input scalar-valued output. Such is made difficult due to sparse imbalanced training data large variations caused both uncertain viewing conditions intrinsic ambiguities observable visual features the scalar values estimated. Encouraged recent success in...

Unsupervised domain adaptation (UDA) is to make predictions for unlabeled data on a target domain, given labeled source whose distribution shifts from the one. Mainstream UDA methods learn aligned features between two domains, such that classifier trained can be readily applied ones. However, transferring strategy has potential risk of damaging intrinsic discrimination data. To alleviate this risk, we are motivated by assumption structural similarity, and propose directly uncover via...

Nonlinear optimization problems with dynamical parameters are widely arising in many practical scientific and engineering applications, various computational models presented for solving them under the hypothesis of short-time invariance. To eliminate large lagging error solution inherently dynamic nonlinear problem, only way is to estimate future unknown information by using present previous data during process, which termed (FDNO) problem. In this paper, suppress noises improve accuracy...

In this paper, the problem of multi-view embedding from different visual cues and modalities is considered. We propose a unified solution for subspace learning methods using Rayleigh quotient, which extensible multiple views, supervised learning, non-linear embeddings. Numerous including Canonical Correlation Analysis, Partial Least Sqaure regression Linear Discriminant Analysis are studied specific intrinsic penalty graphs within same framework. Non-linear extensions based on kernels (deep)...

Semantic patterns offine-grained objects are determined by subtle appearance difference of local parts, which thus inspires a number part-based methods. However, due to uncontrollable object poses in images, distinctive de-tails carried regions can be spatially distributed or even self-occluded, leading large variation on ob-ject representation. For discounting pose variations, this paper proposes learn novel graph based rep-resentation reveal global configuration parts for self-supervised...

In this paper we study automatic recognition of cars four types: Bus, Truck, Van and Small car. For problem consider two data driven frameworks: a deep neural network support vector machine using SIFT features. The accuracy the methods is validated with database over 6500 images, resulting prediction 97 %. This clearly exceeds accuracies earlier studies that use manually engineered feature extraction pipelines.

The point cloud representation of an object can have a large geometric variation in view inconsistent data acquisition procedure, which thus leads to domain discrepancy due diverse and uncontrollable shape cross datasets. To improve discrimination on unseen distribution point-based geometries practical feasible perspective, this paper proposes new method geometry-aware self-training (GAST) for unsupervised adaptation classification. Specifically, aims learn domain-shared semantic categories,...

Fine-grained visual classification can be addressed by deep representation learning under supervision of manually pre-defined targets (e.g., one-hot or the Hadamard codes). Such target coding schemes are less flexible to model inter-class correlation and sensitive sparse imbalanced data distribution as well. In light this, this paper introduces a novel scheme - dynamic relation graphs (DTRG), which, an auxiliary feature regularization, is self-generated structural output mapped from input...

Personalized web search (PWS) has demonstrated its effectiveness in improving the quality of various services on Internet. However, evidences show that users' reluctance to disclose their private information during become a major barrier for wide proliferation PWS. We study privacy protection PWS applications model user preferences as hierarchical profiles. propose framework called UPS can adaptively generalize profiles by queries while respecting user-specified requirements. Our runtime...

We introduce a novel formulation of temporal color constancy which considers multiple frames preceding the frame for illumination is estimated. propose an end-to-end trainable recurrent network – RCC-Net exploits convolutional LSTMs and simulated sequence to learn compositional representations in space time. use standard single benchmark, SFU Gray Ball Dataset, can be adapted setting. Extensive experiments show that proposed method consistently outperforms single-frame state-of-the-art...

This paper seeks to predict the performance of side channel pump by considering influences different wrapping angles. Firstly, three cases 1, 2 and 3 are modeled with angles 15°, 30° 45°, respectively. Secondly, physical parameters comprising exchanged mass flow, pressure velocity distributions plotted at best efficiency point (QBEP) analyze internal flow characteristics. Since exchange times depend on size angle, angle has significant effects head performance. Case 1 smallest recorded...

Abstract Robot manipulators perform a point-point task under kinematic and dynamic constraints. Due to multi-degree-of-freedom coupling characteristics, it is difficult find better desired trajectory. In this paper, multi-objective trajectory planning approach based on an improved elitist non-dominated sorting genetic algorithm (INSGA-II) proposed. Trajectory function planned with new composite polynomial that by combining of quintic polynomials cubic Bezier curves. Then, INSGA-II,...

Recent progress of semantic point clouds analysis is largely driven by synthetic data (e.g., the ModelNet and ShapeNet), which are typically complete, well-aligned noisy-free. Therefore, representations those ideal have limited variations in geometric perspective can gain good performance on a number 3D vision tasks such as cloud classification. In context unsupervised domain adaptation (UDA), representation learning designed for hardly capture invariant patterns from incomplete noisy...

In this study, an efficient hexapod robot interaction system was developed based on the improved digital twin technology. The improves control efficiency and human-computer performance of in complex environments through virtual-real synchronization, adaptive gait planning optimized data transmission virtual model simulates operating state physical entity real time carries out feedback to achieve accurate posture adjustment optimization. Experimental results show that has high synchronization...