Figure–ground segregation: A fully nonlocal approach

Vision Research, Journal Year: 2015, Volume and Issue: 126, P. 308 - 317

Published: April 5, 2015

Language: Английский

---

Relative Depth Order Estimation Using Multi-Scale Densely Connected Convolutional Networks

IEEE Access, Journal Year: 2019, Volume and Issue: 7, P. 38630 - 38643

Published: Jan. 1, 2019

We study the problem of estimating relative depth order point pairs in a monocular image. Recent advances mainly focus on using deep convolutional neural networks to learn and infer ordinal information from multiple contextual pairs, such as global scene context, local information, locations. However, it remains unclear how much each context contributes task. To address this, we first examine contribution cue performance estimation. find out that surrounding most, helps little. Based findings, propose simple method, multi-scale densely-connected network tackle Instead learning structure, dedicate explore structure by regress regions sizes around pairs. Moreover, use recent densely connected encourage substantial feature reuse well deepen our boost performance. show experiments results approach are par with or better than state-of-the-art methods benefit only small number training data.

Language: Английский

---

3D Depth Perception from Single Monocular Images

Lecture notes in computer science, Journal Year: 2014, Volume and Issue: unknown, P. 510 - 521

Published: Dec. 22, 2014

Language: Английский

---

Monocular Depth Ordering using Perceptual Occlusion Cues

Published: Jan. 1, 2016

In this paper we propose a method to estimate global depth order between the objects of scene using information from single image coming an uncalibrated camera.The present stems early vision cues such as occlusion and convexity uses them infer both local order.Monocular cues, namely, T-junctions convexities, contain suggesting neighbouring objects.A combination these is more suitable, because, while conveyed by perceptually stronger, they are not prevalent in natural images.We novel detector that also establishes order.The partial extracted curvature-based multi-scale feature.Finally, order, i.e., full all shapes consistent possible with computed orders can tolerate conflicting computed.An integration scheme based on Markov chain approximation rank aggregation problem used for purpose.The experiments conducted show proposed compares favorably state art.

Language: Английский

---

Neighborhood Filters and the Recovery of 3D Information

Springer eBooks, Journal Year: 2015, Volume and Issue: unknown, P. 1645 - 1673

Published: Jan. 1, 2015

Language: Английский

---

Image Cartoon-Texture Decomposition Using Isotropic Patch Recurrence

arXiv (Cornell University), Journal Year: 2018, Volume and Issue: unknown

Published: Jan. 1, 2018

Aiming at separating the cartoon and texture layers from an image, cartoon-texture decomposition approaches resort to image priors model respectively. In recent years, patch recurrence has emerged as a powerful prior for recovery. However, existing strategies of using are ineffective decomposition, both contours patterns exhibit strong in images. To address this issue, we introduce isotropy recurrence, that spatial configuration similar patches exhibits isotropic structure which is different cartoon, component. Based on construct nonlocal sparsification system can effectively distinguish well-patterned features contour edges. Incorporating constructed into morphology component analysis, develop effective method noiseless noisy decomposition. The experimental results have demonstrated superior performance proposed ones, well effectiveness prior.

Language: Английский

---

Monocular depth estimation in images and sequences using occlusion cues

Published: Feb. 21, 2014

When humans observe a scene, they are able to perfectly distinguish the different parts composing it. Moreover, can easily reconstruct spatial position of these and conceive consistent structure. The mechanisms involving visual perception have been studied since beginning neuroscience but, still today, not all processes it known. In usual situations, make use three methods estimate scene first one is so called divergence makes both eyes. objects lie in front observed at distance up hundred meters, subtle differences image formation each eye be used determine depth. field view eyes, other should used. cases, cues prior learned information Even if less accurate than divergence, almost always infer correct depth structure when using them. As an example cues, occlusion, perspective or object size provide lot about scene. A priori depends on observer, but normally subconsciously by detect commonly known regions such as sky, ground types objects. last years, technology has handle processing burden vision systems, there lots efforts devoted design automated interpreting systems. this thesis we address problem estimation only point occlusion cues. objective occlusions present combine them with segmentation system generate relative order map for We explore static dynamic situations single images, frame inside sequences full video sequences. case where sequence available, exploiting motion recover also designed. Results promising competitive respect state art literature, much room improvement compared human performance. Quan els observen una escena, son capaços de distingir perfectament les que la composen i organitzar-les espacialment per tal poder-se orientar. Els mecanismes governen percepció han estat estudiats des dels principis neurociència, però encara no es coneixen tots processos biològic hi prenen part. En situacions normals, poden fer servir tres eines estimar l’estructura l’escena. La primera és l’anomenada divergència. Aprofita l’ús dos punts vista (els ulls) capaç¸ determinar molt acuradament posició objectes ,que distància fins cent metres, romanen enfront l’observador. mesura augmenta o troben en el camp visió ulls, altres s’han d’utilitzar. Tant l’experiència anterior com certs indicis visuals s’utilitzen aquests casos i, seva precisió menor, aconsegueixen quasi bé sempre interpretar seu entorn. aporten informació profunditat més coneguts utilitzats són exemple, perspectiva, oclusions tamany objectes. L’experiència permet resoldre vistes anteriorment ara saber quins corresponen al terra, cel Durant últims anys, quan tecnologia ho ha permès, intentat dissenyar sistemes interpretessin automàticament diferents tipus d’escena. aquesta tesi s’aborda tema l’estimació utilitzant només un punt d’oclusió. L’objectiu del treball detecció d’aquests combinar-los amb sistema segmentació generar plans presents escena. explora tant estàtiques (imatges fixes) dinàmiques, trames dins seqüències vídeo completes. cas completes, també proposa automàtic reconstruir l’escena moviment. resultats prometedors competitius literatura moment, mostren computador té marge millora respecte humans.

---

Emergence of border-ownership by large-scale consistency and long-range interactions: Neuro-computational model to reflect global configurations.

Psychological Review, Journal Year: 2021, Volume and Issue: 128(4), P. 597 - 622

Published: June 3, 2021

The visual system performs remarkably well to perceive depth order of surfaces without stereo disparity, indicating the importance figure-ground organization based on pictorial cues. To understand how emerges, it is essential investigate global configuration an image reflected. In past, many neuro-computational models developed reproduce implemented algorithms give a bias convex areas. However, in certain conditions, area can be perceived as hole and nonconvex figural. This occurs when surface properties are consistent with background and, hence, grouped together our perception. We argue that large-scale consistency reflected border-ownership computation. model, called DISC2, first analyzes relationships between two signals all possible combinations image. It then enhances if they satisfy following conditions: (a) fit (b) at locations consistent. strength enhancement decays distance signals. model gives extremely robust responses various images complexities both shape order. Furthermore, we advanced version ("augmented model") where computation above interacts local curvilinearity, which further enhanced nature model. results suggest involvement similar computational processes brain for organization. (PsycInfo Database Record (c) 2021 APA, rights reserved).

Language: Английский

---

From Occlusion to Global Depth Order, a Monocular Approach

Communications in computer and information science, Journal Year: 2017, Volume and Issue: unknown, P. 575 - 592

Published: Jan. 1, 2017

Language: Английский

---

A deep generative directed network for scene depth ordering

Journal of Visual Communication and Image Representation, Journal Year: 2018, Volume and Issue: 58, P. 554 - 564

Published: Dec. 18, 2018

Language: Английский

---