Affinity CNN: Learning Pixel-Centric Pairwise Relations for Figure/Ground Embedding

Published: June 1, 2016

Spectral embedding provides a framework for solving perceptual organization problems, including image segmentation and figure/ground organization. From an affinity matrix describing pairwise relationships between pixels, it clusters pixels into regions, and, using complex-valued extension, orders according to layer. We train convolutional neural network (CNN) directly predict the pair-wise that define this matrix. then resolves these predictions globally-consistent of scene. Experiments demonstrate significant benefit direct coupling compared prior works which use explicit intermediate stages, such as edge detection, on pathway from affinities. Our results suggest spectral powerful alternative conditional random field (CRF)-based globalization schemes typically coupled deep networks.

Language: Английский

---

Mining health knowledge graph for health risk prediction

World Wide Web, Journal Year: 2020, Volume and Issue: 23(4), P. 2341 - 2362

Published: March 20, 2020

Language: Английский

---

DOC: Deep OCclusion Estimation from a Single Image

Lecture notes in computer science, Journal Year: 2016, Volume and Issue: unknown, P. 545 - 561

Published: Jan. 1, 2016

Language: Английский

---

Stacked ensemble modeling for improved tuberculosis treatment outcome prediction in pediatric cases

Concurrency and Computation Practice and Experience, Journal Year: 2024, Volume and Issue: 36(13)

Published: March 17, 2024

Summary The promising results of ML (machine learning) methods in various disciplines have led to the frequent use these health fields such as disease diagnosis, personalized medicine, medical image‐based and predicting number deaths cases a pandemic. However, neglected area field healthcare is lack study with predict treatment outcomes for tuberculosis (TB) patients, particularly children experiencing failed treatment. This need has become more apparent coronavirus pandemic reversed gains institutions TB disease, especially children. Therefore, this article conducted using stacked ensemble method early risk outcome TB. To fulfill determine most appropriate technique, two‐stage methodology was followed work. First, predictions were obtained by combining information gain feature selection (IGFS) approach variety single‐based algorithms, including logistic regression (LR), deep belief neural networks (DBN), random forest (RF), decision tree (DT). Second, proposed method, which includes used. latter model uses LR meta‐learner aforementioned algorithms (DBN, LR, RF, DT). performance models used two stages compared, combination stack‐based learning IGFS technique provided better ROC curves, accuracy, precision, recall results.

Language: Английский

---

Integrating Both Parallax and Latency Compensation into Video See-through Head-mounted Display

IEEE Transactions on Visualization and Computer Graphics, Journal Year: 2023, Volume and Issue: 29(5), P. 2826 - 2836

Published: Feb. 27, 2023

This work introduces a perspective-corrected video see-through mixed-reality head-mounted display with edge-preserving occlusion and low-latency capabilities. To realize the consistent spatial temporal composition of captured real world containing virtual objects, we perform three essential tasks: 1) to reconstruct images so as match user's view; 2) occlude objects nearer provide users correct depth cues; 3) reproject scenes be matched keep up users' head motions. Captured image reconstruction occlusion-mask generation require dense accurate maps. However, estimating these maps is computationally difficult, which results in longer latencies. obtain an acceptable balance between consistency low latency, rapidly generated by focusing on edge smoothness disocclusion (instead fully maps), shorten processing time. Our algorithm refines edges via hybrid method involving infrared masks color-guided filters, it fills disocclusions using temporally cached system combines algorithms two-phase warping architecture based upon synchronized camera pairs displays. The first phase reduce registration errors scenes. second present that correspond motion. We implemented methods our wearable prototype performed end-to-end measurements its accuracy latency. achieved latency due motion (less than 4 ms) 0.1° size less 0.3° position) test environment. anticipate this will help improve realism mixed reality systems.

Language: Английский

---

Mining Heterogeneous Information Graph for Health Status Classification

Published: Nov. 1, 2018

In the medical domain, there exists a large volume of data from multiple sources such as electronic health records, general examination results and surveys. The contain useful information reflecting people's provides great opportunities for studies to improve quality healthcare. However, how mine these effectively efficiently still remains critical challenge. this paper, we propose an innovative classification model knowledge discovery patients' personal repositories. By based on analytics massive in National Health Nutrition Examination Survey, study builds classify status reveal specific disease potentially suffered by patient. This paper makes significant contributions advancement mining with specifically crafted domain-based data. Moreover, research contributes healthcare community providing deep understanding accessibility patterns various observations.

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

---

CONTEMPORARY DATA ANALYSIS TECHNIQUES FOR ONLINE REPUTATION MANAGEMENT IN HOSPITALITY AND TOURISM

Facta Universitatis Series Economics and Organization, Journal Year: 2019, Volume and Issue: unknown, P. 059 - 059

Published: May 28, 2019

Knowing what attracts or deters tourists to/from a tourist visit and products to offer them pay special attention is crucial for good economic results. Such knowledge can be obtained by analysis of online comments reviews that leave on travel websites (such as Booking, TripAdvisor, Trivago, etc.). This paper describes the value which information about opinions emotions hidden in has managers who receive it, especially (dis)satisfaction users with certain aspects offer. Uncovered from provides chance take advantage strong points, correct shortcomings through timely corrective measures actions. Contemporary approaches methods analyzing opportunities development they provide tourism industry are described case study conducted over subset 20491 hotel TripAdvisor. We have sentiment goal building an automated model will successfully distinguish positive negative reviews. Logistic Regression classifier best performance, 90% it correctly classified 83% negative. illustrated how association rules help management uncover relationships between concepts under discussion

Language: Английский

---

Study of Manhattan and Region Growing Methods for Brain Tumor Detection

Journal of Advances in Information Technology, Journal Year: 2024, Volume and Issue: 15(2), P. 183 - 194

Published: Jan. 1, 2024

This paper investigates the utilization of regiongrowing segmentation and Content-Based Image Retrieval (CBIR) techniques to predict brain cancer, particularly focusing on tumors.Recent advancements in medical science have brought about promising diagnostic methods treatments, offering patients renewed hope for recovery.However, existing problems diagnosing cancer include time inefficiency, inconsistency, inaccuracy, costly.Hence, this study aims find an innovative approach address predicaments diagnosis by harnessing power artificial intelligence, specifically within realm computer vision.The CBIR are employed purpose.To presence tumors, these applied CT-scan images.The dataset comprises over 800 images sourced from Kaggle.com a hospital Lampung, Indonesia.The effectiveness region-growing method is evaluated using Receiver Operating Characteristics (ROC) analysis, along with assessment quality affected regions demonstrates that achieve accuracy rate 79% when tested consisting 400 normal images.Simultaneously, image retrieval remarkable, surpassing 96% 94% Manhattan Euclidean distance metrics, respectively.In conclusion, findings research indicate combination can substantially enhance performance algorithms designed tumor detection.

Language: Английский

---

Motion Occlusions for Automatic Generation of Relative Depth Maps

Published: Sept. 7, 2018

Recovering of the depth structure a scene from monocular video content provides an important advantage in applications such as AR (placing and removing objects) or 3D-TV 3D cinema (2D-to-3D conversion). In this paper, we present automatic method to generate relative maps sequences. It relies on dynamic occlusion cue recover order objects scene. The forward backward motion analysis between each two consecutive frames allows calculation their occlusions. We estimate using modified version EpicFlow. Our modifications optical flow made it coherent forward-backward directions without compromising its performance. Thanks new feature, occlusions are simpler calculate than approaches used relevant literature. obtained allow deduction contained image. These segmentation approach which considers both color motion. Ours results show small improvement quality while adding forward/backward coherence. With respect ordering our obtains slightly better reference computationally costly step processing.

Language: Английский

---