This paper proposes a novel joint learning and densely-cooperative fusion (JL-DCF) architecture for RGB-D salient object detection. Existing models usually treat RGB depth as independent information design separate networks feature extraction from each. Such schemes can easily be constrained by limited amount of training data or over-reliance on an elaborately-designed process. In contrast, our JL-DCF learns both inputs through Siamese network. To this end, we propose two effective...

Modern crowd counting methods usually employ deep neural networks (DNN) to estimate counts via density regression. Despite their significant improvements, the regression-based are incapable of providing detection individuals in crowds. The detection-based methods, on other hand, have not been largely explored recent trends due needs for expensive bounding box annotations. In this work, we instead propose a new network with only point supervision required. It can simultaneously detect size...

Existing RGB-D salient object detection (SOD) models usually treat RGB and depth as independent information design separate networks for feature extraction from each. Such schemes can easily be constrained by a limited amount of training data or over-reliance on an elaborately designed process. Inspired the observation that modalities actually present certain commonality in distinguishing objects, novel joint learning densely cooperative fusion (JL-DCF) architecture is to learn both inputs...

RGB-D salient object detection (SOD) recently has attracted increasing research interest and many deep learning methods based on encoder-decoder architectures have emerged. However, most existing SOD models conduct feature fusion either in the single encoder or decoder stage, which hardly guarantees sufficient cross-modal ability. In this paper, we make first attempt addressing through 3D convolutional neural networks. The proposed model, named RD3D, aims at pre-fusion stage in-depth to...

This paper proposes an encoder-decoder network to disentangle shape features during 3D face reconstruction from single 2D images, such that the tasks of reconstructing accurate shapes and learning discriminative for recognition can be accomplished simultaneously. Unlike existing methods, our proposed method directly regresses dense tackles identity residual (i.e., non-identity) components in explicitly separately based on a composite model with latent representations. We devise training...

Face alignment and 3D face reconstruction are traditionally accomplished as separated tasks. By exploring the strong correlation between 2D landmarks shapes, in contrast, we propose a joint method to simultaneously solve these two problems for images of arbitrary poses expressions. This method, based on summation model faces cascaded regression shape spaces, iteratively alternately applies regressors, one updating other shape. The correlated via 3D-to-2D mapping matrix, which is updated each...

RGB-D salient object detection (SOD) recently has attracted increasing research interest by benefiting conventional RGB SOD with extra depth information. However, existing models often fail to perform well in terms of both efficiency and accuracy, which hinders their potential applications on mobile devices real-world problems. An underlying challenge is that the model accuracy usually degrades when simplified have few parameters. To tackle this dilemma also inspired fact quality a key...

Depth information has been proved beneficial in RGB-D salient object detection (SOD). However, depth maps obtained often suffer from low quality and inaccuracy. Most existing SOD models have no cross-modal interactions or only unidirectional to RGB their encoder stages, which may lead inaccurate features when facing depth. To address this limitation, we propose conduct progressive bidirectional as early the stage, yielding a novel bi-directional transfer-and-selection network named BTS-Net,...

With the increasing demand for search and rescue, it is highly demanded to detect objects of interest in large-scale images captured by Unmanned Aerial Vehicles (UAVs), which quite challenging due extremely small scales objects. Most existing methods employed Feature Pyramid Network (FPN) enrich shallow layers' features combing deep contextual features. However, under limitation inconsistency gradient computation across different layers, layers FPN are not fully exploited tiny In this paper,...

RGB-depth (RGB-D) salient object detection (SOD) recently has attracted increasing research interest, and many deep learning methods based on encoder–decoder architectures have emerged. However, most existing RGB-D SOD models conduct explicit controllable cross-modal feature fusion either in the single encoder or decoder stage, which hardly guarantees sufficient ability. To this end, we make first attempt addressing through 3-D convolutional neural networks. The proposed model, named RD3D,...

High-resolution automated fingerprint recognition systems (AFRSs) offer higher security because they are able to make use of level-3 features, such as pores, that not available in lower resolution ( <; 500-dpi) images. One the main parameters affecting quality a digital image and issues cost, interoperability, performance an AFRS is choice resolution. In this paper, we identify optimal for using two most representative features: minutiae pores. We first designed multiresolution acquisition...

Compared to conventional saliency detection by handcrafted features, deep convolutional neural networks (CNNs) recently have been successfully applied field with superior performance on locating salient objects. However, due repeated sub-sampling operations inside CNNs such as pooling and convolution, many CNN-based models fail maintain fine-grained spatial details boundary structures of To remedy this issue, paper proposes a novel end-to-end learning-based refinement model named Refinet,...

Estimating homography from an image pair is a fundamental problem in alignment. Unsupervised learning methods have received increasing attention this field due to their promising performance and label-free training. However, existing do not explicitly consider the of plane-induced parallax, which will make predicted compromised on multiple planes. In work, we propose novel method HomoGAN guide unsupervised estimation focus dominant plane. First, multi-scale transformer network designed...