SIFT-like local feature descriptors are ubiquitously employed in computer vision applications such as content-based retrieval, video analysis, copy detection, object recognition, photo tourism, and 3D reconstruction. Feature can be designed to invariant certain classes of photometric geometric transformations, particular, affine intensity scale transformations. However, real transformations that an image undergo only approximately modeled this way, thus most practice. Second, usually high...

Distance metric learning (DML) has been successfully applied to object classification, both in the standard regime of rich training data and few-shot scenario, where each category is represented by only a few examples. In this work, we propose new method for DML that simultaneously learns backbone network parameters, embedding space, multi-modal distribution categories single end-to-end process. Our approach outperforms state-of-the-art methods DML-based classification on variety...

We introduce a new framework for learning dense correspondence between deformable 3D shapes. Existing based approaches model shape as labelling problem, where each point of query receives label identifying on some reference domain; the is then constructed posteriori by composing predictions two input propose paradigm shift and design structured prediction in space functional maps, linear operators that provide compact representation correspondence. process via deep residual network which...

The availability of affordable and portable depth sensors has made scanning objects people simpler than ever. However, dealing with occlusions missing parts is still a significant challenge. problem reconstructing (possibly non-rigidly moving) 3D object from single or multiple partial scans received increasing attention in recent years. In this work, we propose novel learning-based method for the completion shapes. Unlike majority existing approaches, our focuses on that can undergo...

We present DeepISP, a full end-to-end deep neural model of the camera image signal processing (ISP) pipeline. Our learns mapping from raw low-light mosaiced to final visually compelling and encompasses low-level tasks such as demosaicing denoising well higher-level color correction adjustment. The training evaluation pipeline were performed on dedicated dataset containing pairs well-lit images captured by Samsung S7 smartphone in both processed JPEG formats. proposed solution achieves...

Learning to classify new categories based on just one or a few examples is long-standing challenge in modern computer vision. In this work, we proposes simple yet effective method for few-shot (and one-shot) object recognition. Our approach modified auto-encoder, denoted Delta-encoder, that learns synthesize samples an unseen category by seeing from it. The synthesized are then used train classifier. proposed both extract transferable intra-class deformations, "deltas", between same-class...

We introduce the first completely unsupervised correspondence learning approach for deformable 3D shapes. Key to our model is understanding that natural deformations (such as changes in pose) approximately preserve metric structure of surface, yielding a criterion drive process toward distortion-minimizing predictions. On this basis, we overcome need annotated data and replace it by purely geometric criterion. The resulting class-agnostic, able leverage any type training phase. In contrast...

Informative and discriminative feature descriptors play a fundamental role in deformable shape analysis. For example, they have been successfully employed correspondence, registration, retrieval tasks. In recent years, significant attention has devoted to obtained from the spectral decomposition of Laplace-Beltrami operator associated with shape. Notable examples this family are heat kernel signature (HKS) recently introduced wave (WKS). The Laplacian-based achieve state-of-the-art...

Three important properties of a classification machinery are: (i) the system preserves core information input data; (ii) training examples convey about unseen and (iii) is able to treat differently points from different classes. In this work we show that these fundamental are satisfied by architecture deep neural networks. We formally prove networks with random Gaussian weights perform distance-preserving embedding data, special treatment for in-class out-of-class data. Similar at network...

Parsimony, including sparsity and low rank, has been shown to successfully model data in numerous machine learning signal processing tasks. Traditionally, such modeling approaches rely on an iterative algorithm that minimizes objective function with parsimony-promoting terms. The inherently sequential structure data-dependent complexity latency of optimization constitute a major limitation many applications requiring real-time performance or involving large-scale data. Another encountered by...

Abstract The recent introduction of 3D shape analysis frameworks able to quantify the deformation a into another in terms variation real functions yields new interpretation similarity assessment and opens perspectives. Indeed, while classical approaches mainly it as numerical score, map‐based methods also define (dense) correspondences. After presenting detail theoretical foundations underlying these approaches, we classify them by looking at their most salient features, including kind...

Depth estimation from a single image is well-known challenge in computer vision. With the advent of deep learning, several approaches for monocular depth have been proposed, all which inherent limitations due to scarce cues that exist image. Moreover, these methods are very demanding computationally, makes them inadequate systems with limited processing power. In this paper, phase-coded aperture camera proposed. The equipped an optical phase mask provides unambiguous depth-related color...

Many algorithms for the computation of correspondences between deformable shapes rely on some variant nearest neighbor matching in a descriptor space. Such are, example, various point-wise correspondence recovery used as post-processing stage functional framework. frequently techniques implicitly make restrictive assumptions (e.g., nearisometry) considered and practice suffer from lack accuracy result poor surjectivity. We propose an alternative technique capable guaranteeing bijective...

Example synthesis is one of the leading methods to tackle problem few-shot learning, where only a small number samples per class are available. However, current approaches address scenario single category label image. In this work, we propose novel technique for synthesizing with multiple labels (yet unhandled) multi-label classification scenario. We combine pairs given examples in feature space, so that resulting synthesized vectors will correspond whose sets obtained through certain set...

We propose an efficient procedure for calculating partial dense intrinsic correspondence between deformable shapes performed entirely in the spectral domain. Our technique relies on recently introduced functional maps formalism and joint approximate diagonalization (JAD) of Laplace-Beltrami operators previously matching non-isometric shapes. show that a variant JAD problem with appropriately modified coupling term (surprisingly) allows to construct quasi-harmonic bases localized latent...

We present a method to match three dimensional shapes under non-isometric deformations, topology changes and partiality. formulate the problem as matching between set of pair-wise point-wise descriptors, imposing continuity prior on mapping, propose projected descent optimization procedure inspired by difference convex functions (DC) programming.

Abstract We present a method for supervised learning of shape descriptors retrieval applications. Many content‐based approaches follow the bag‐of‐features (BoF) paradigm commonly used in text and image by first computing local descriptors, then representing them ‘geometric dictionary’ using vector quantization. A major drawback such is that dictionary constructed an unsupervised manner clustering, unaware last stage process (pooling into BoF, comparison latter some metric). In this paper, we...

We propose a fully convolutional neural-network architecture for image denoising which is simple yet powerful. Its structure allows to exploit the gradual nature of process, in shallow layers handle local noise statistics, while deeper recover edges and enhance textures. Our method advances state art when trained different levels distributions (both Gaussian Poisson). In addition, we show that making denoiser class-aware by exploiting semantic class information boosts performance, enhances...

The success of learning with noisy labels (LNL) methods relies heavily on the a warm-up stage where standard supervised training is performed using full (noisy) set. In this paper, we identify "warm-up obstacle": inability stages to train high quality feature extractors and avert memorization labels. We propose "Contrast Divide" (C2D), simple framework that solves problem by pre-training extractor in self-supervised fashion. Using boosts performance existing LNL approaches drastically...

Detection and description of affine-invariant features is a cornerstone component in numerous computer vision applications. In this note, we analyze the notion maximally stable extremal regions (MSERs) through prism curvature scale space, conclude that its original definition, MSER prefers regular (round) regions. Arguing interesting natural images usually have irregular shapes, propose alternative definitions which are free bias, yet maintain their invariance properties.

We consider the problem of minimum distortion intrinsic correspondence between deformable shapes, many useful formulations which give rise to NP-hard quadratic assignment (QAP). Previous attempts use spectral relaxation have had limited success due lack sparsity obtained "fuzzy" solution. In this paper, we adopt recently introduced alternative L <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> QAP based on principles game theory. relate...

Comparison among graphs is ubiquitous in graph analytics. However, it a hard task terms of the expressiveness employed similarity measure and efficiency its computation. Ideally, comparison should be invariant to order nodes sizes compared graphs, adaptive scale patterns, scalable. Unfortunately, these properties have not been addressed together. Graph comparisons still rely on direct approaches, kernels, or representation-based methods, which are all inefficient impractical for large...