The research focus of designing local patch descriptors has gradually shifted from handcrafted ones (e.g., SIFT) to learned ones. In this paper, we propose learn high performance descriptor in Euclidean space via the Convolutional Neural Network (CNN). Our method is distinctive four aspects: (i) We a progressive sampling strategy which enables network access billions training samples few epochs. (ii) Derived basic concept matching problem, empha-size relative distance between descriptors....

Current state-of-the-art detectors typically exploit feature pyramid to detect objects at different scales. Among them, FPN is one of the representative works that build a by multi-scale features summation. However, design defects behind prevent from being fully exploited. In this paper, we begin first analyzing in FPN, and then introduce new architecture named AugFPN address these problems. Specifically, consists three components: Consistent Supervision, Residual Feature Augmentation, Soft...

Despite the fact that Second Order Similarity (SOS) has been used with significant success in tasks such as graph matching and clustering, it not exploited for learning local descriptors. In this work, we explore potential of \sos field descriptor by building upon intuition a positive pair points should exhibit similar distances respect to other embedding space. Thus, propose novel regularization term, named Regularization (SOSR), follows principle. By incorporating SOSR into training, our...

This paper presents a novel method for feature description based on intensity order. Specifically, Local Intensity Order Pattern(LIOP) is proposed to encode the local ordinal information of each pixel and overall used divide patch into subregions which are accumulating LIOPs respectively. Therefore, both captured by LIOP descriptor so as make it highly discriminative descriptor. It shown that not only invariant monotonic changes image rotation but also robust many other geometric photometric...

Hyperspectral unmixing, the process of estimating a common set spectral bases and their corresponding composite percentages at each pixel, is an important task for hyperspectral analysis, visualization understanding. From unsupervised learning perspective, this problem very challenging---both are unknown, making solution space too large. To reduce space, many approaches have been proposed by exploiting various priors. In practice, these priors would easily lead to some unsuitable solution....

Although traffic sign recognition has been studied for many years, most existing works are focused on the symbol-based signs. This paper proposes a new data-driven system to recognize all categories of signs, which include both and text-based in video sequences captured by camera mounted car. The consists three stages, regions interest (ROIs) extraction, ROIs refinement classification, post-processing. Traffic from each frame first extracted using maximally stable extremal gray normalized...

This paper proposes a novel method for interest region description which pools local features based on their intensity orders in multiple support regions. Pooling by is not only invariant to rotation and monotonic changes, but also encodes ordinal information into descriptor. Two kinds of are used this paper, one gradients the other intensities; hence, two descriptors obtained: Multisupport Region Order-Based Gradient Histogram (MROGH) Rotation Intensity Monotonic Invariant Descriptor...

Although feature-based methods have been successfully developed in the past decades for registration of optical images, and synthetic aperture radar (SAR) images is still a challenging problem remote sensing. In this letter, an improved version scale-invariant feature transform first proposed to obtain initial matching features from SAR images. Then, are refined by exploring their spatial relationship. The matches finally used estimating parameters. Experimental results shown effectiveness method.

A novel local image descriptor is proposed in this paper, which combines intensity orders and gradient distributions multiple support regions. The novelty lies three aspects: 1) calculated a rotation invariant way given region; 2) gradients are adaptively pooled spatially based on order to encode spatial information; 3) Multiple regions used for constructing further improves its discriminative ability. Therefore, the encodes not only information but also about relative relationship of...

Pixel-level classification for very high resolution (VHR) images is a crucial but challenging task in remote sensing. However, since the diverse ways of satellite image acquisition and distinct structures various regions, distributions same semantic classes among different data sets are dissimilar. Therefore, model trained on one set (source domain) may collapse, when it directly applied to another (target domain). To solve this problem, many adversarial-based domain adaptation methods have...

Scene classification is one of the most important issues in remote sensing image processing. To obtain a high discriminative feature representation for an to be classified, traditional methods usually consider densely accumulate hand-crafted low-level descriptors (e.g., scale-invariant transform) by encoding techniques. However, performance largely limited as they are not capable describing rich semantic information contained various images. alleviate this problem, we propose novel method...

Transferring image-based object detectors to the domain of videos remains a challenging problem. Previous efforts mostly exploit optical flow propagate features across frames, aiming achieve good trade-off between accuracy and efficiency. However, introducing an extra model estimate can significantly increase overall size. The gap high-level also hinder it from establishing spatial correspondence accurately. Instead relying on flow, this paper proposes novel module called Progressive Sparse...

This paper performs a comprehensive and comparative evaluation of the state-of-the-art local features for task image-based 3D reconstruction. The evaluated cover recently developed ones by using powerful machine learning techniques elaborately designed handcrafted features. To obtain evaluation, we choose to include both float type binary ones. Meanwhile, two kinds datasets have been used in this evaluation. One is dataset many different scene types with groundtruth points, containing images...

Nowdays, it is prevalent to train deep learning (DL) models in cloud-native platforms that actively leverage containerization and orchestration technologies for high elasticity, low flexible operation cost, many other benefits. However, also faces new challenges our work focusing on those related I/O throughput training, including complex data access with complicated performance tuning, lack of cache capacity specialized hardware match its dynamic requirement, inefficient resource sharing...

Open-set action recognition (OSAR) aims to learn a framework capable of both classifying known classes and identifying unknown actions in open-set scenarios. Existing OSAR methods typically reside data-driven paradigm, which ignore the rich semantics categories. In fact, we humans have capability leveraging captured semantic information, i.e., knowledge experience, incisively distinguish samples from classes. Motivated by this observation, paper, propose Unified Semantic Exploration (USE)...

Object detection accuracy degrades seriously in visually degraded scenes. A natural solution is to first enhance the image and then perform object detection. However, it suboptimal does not necessarily lead improvement of due separation enhancement tasks. To solve this problem, we propose an guided method, which refines network with additional branch end-to-end way. Specifically, are organized a parallel way, feature module designed connect two branches, optimizes shallow input be as...

Underwater image enhancement aims to restore a clean appearance and thus improves the quality of underwater degraded images. Current methods feed whole directly into model for enhancement. However, they ignored that R, G B channels images present varied degrees degradation, due selective absorption light. To address this issue, we propose an unsupervised multi-expert learning by considering each color channel. Specifically, architecture based on generative adversarial network is employed...

A novel method for line matching is proposed. The basic idea to use tentative point correspondences, which can be easily obtained by keypoint methods, significantly improve performance, even when the correspondences are severely contaminated outliers. When a pair of image lines, group corresponding points that may coplanar with these lines in 3D space firstly from all local neighborhoods lines. Then given such points, similarity between this calculated based on an affine invariant one and...

Feature description for local image patch is widely used in computer vision. While the conventional way to design descriptor based on expert experience and knowledge, learning-based methods designing become more popular because of their good performance data-driven property. This paper proposes a novel method binary feature descriptor, which we call receptive fields (RFD). Technically, RFD constructed by thresholding responses set fields, are selected from large number candidates according...

We propose to integrate spectral-spatial feature extraction and tensor discriminant analysis for hyperspectral image classification. First, we apply remarkable approaches in the cube extract a each pixel. Then, based on class label information, local is used remove redundant information subsequent classification procedure. The approach not only extracts sufficient features from original images but also gets better representation owing framework. Comparative results two benchmarks demonstrate...

More attention has been given to studies on the synchronous enhancement of light-to-thermal conversion capacity and thermal conductivity polymer phase-change material (PCM) conserve solar energy. Here, commonly used PEG PCM was encapsulated by graphene nanoplatelets (GNPs) single-walled carbon nanotubes (SWCNs), while composites were simultaneously obtained. The 3D interconnected SWCNs GNPs equipped with (1) shape stability durability, (2) negligible change in energy storage density, (3)...