1,720,972 research outputs found
AerialWaste dataset for landfill discovery in aerial and satellite images
Get PDFIllegal landfills are sites where garbage is dumped violating waste management laws. Aerial images enable the use of photo interpretation for territory scanning and landfill detection but this practice is hindered by the manual nature of this task which also requires expert knowledge. Deep Learning methods can help capture the analysts’ expertise and build automated landfill discovery tools. However, this goal requires public high-quality datasets for model training and testing. At present no such datasets exist and this gap penalizes the research toward scalable and accurate landfill discovery methods. We present a dataset for landfill detection featuring airborne, WorldView-3, and GoogleEarth images annotated by professional photo interpreters. It comprises 3,478 positive and 6,956 negative examples. Most positive instances are characterized by metadata: the type of waste, its storage mode, the type of the site, and the evidence and severity of the illicit. The dataset has been technically validated by building an accurate landfill detector and is accompanied by a visualization and annotation tool
Heterogeneous information integration for mountain augmented reality mobile apps
Get PDFMobile Augmented Reality (AR) applications offer a new way to promote the collection of geo-referenced information, by engaging citizens in a useful experience and encouraging them to gather environment data, such as images of plant species or of mountain snow coverage. The distinctive characteristic of mobile AR applications is the overlay of information directly on top of what the user sees, based on the user’s context estimated from the device sensors. The application analyzes the sensor readings (GPS position, phone orientation and motion, and possibly also the camera frame content), to understand what the user is watching and enriches the view with contextual information. Developing mobile AR applications poses several challenges related to the acquisition, selection, transmission and display of information, which gets more demanding in mountain applications where usage without Internet connectivity is a strong requirement. This paper discusses the experience of a real world mobile AR application for mountain exploration, which can be used to crowdsource the collection of mountain images for environmental purposes, such as the analysis of snow coverage for water availability prediction and the monitoring of plant diseases
Convolutional neural network for pixel-wise skyline detection
Get PDFOutdoor augmented reality applications are an emerging class of software systems that demand the fast identification of natural objects, such as plant species or mountain peaks, in low power mobile devices. Convolutional Neural Networks (CNN) have exhibited superior performance in a variety of computer vision tasks, but their training is a labor intensive task and their execution requires non negligible memory and CPU resources. This paper presents the results of training a CNN for the fast extraction of mountain skylines, which exhibits a good balance between accuracy (94,45% in best conditions and 86,87% in worst conditions), memory consumption (9,36 MB on average) and runtime execution overhead (273 ms on a Nexus 6 mobile phone), and thus has been exploited for implementing a real-world augmented reality applications for mountain peak recognition running on low to mid-end mobile phones
ALMOsT-trace: A web based embeddable tracing tool for ALMOsT.js
No full textModel Driven Development (MDD) requires model-to-model and/or model-to-text transformations to produce application code from high level descriptions. Debugging and evaluating such transformations is in itself a complex task; complexity which can be mitigated through the usage of advanced developer tools. We demonstrate ALMOsT-Trace, a plug-in for ALMOsT.js, which allows developers to debug and analyze their model transformations from within their applications. In the demo, attendees will be able to experiment with ALMOsT-Trace by evaluating it in IFMLEdit.org, an online tool for the rapid prototyping of web and mobile applications, and by means of examples that can be customized by the attendees themself
Black-box Error Diagnosis in Deep Neural Networks for Computer Vision: a Survey of Tools
Get PDFThe application of Deep Neural Networks (DNNs) to a broad variety of tasks
demands methods for coping with the complex and opaque nature of these
architectures. When a gold standard is available, performance assessment treats
the DNN as a black box and computes standard metrics based on the comparison of
the predictions with the ground truth. A deeper understanding of performances
requires going beyond such evaluation metrics to diagnose the model behavior
and the prediction errors. This goal can be pursued in two complementary ways.
On one side, model interpretation techniques "open the box" and assess the
relationship between the input, the inner layers and the output, so as to
identify the architecture modules most likely to cause the performance loss. On
the other hand, black-box error diagnosis techniques study the correlation
between the model response and some properties of the input not used for
training, so as to identify the features of the inputs that make the model
fail. Both approaches give hints on how to improve the architecture and/or the
training process. This paper focuses on the application of DNNs to Computer
Vision (CV) tasks and presents a survey of the tools that support the black-box
performance diagnosis paradigm. It illustrates the features and gaps of the
current proposals, discusses the relevant research directions and provides a
brief overview of the diagnosis tools in sectors other than CV.Comment: Accepted in Springer Neural Computing and Application
Automatic feature extraction to support Mountains Mapping in OSM
No full textNahime Torres et al. (2019). Automatic feature extraction to support Mountains Mapping in OSM
In: Minghini, M., Grinberger, A.Y., Juhász, L., Yeboah, G., Mooney, P. (Eds.). Proceedings of the Academic Track at the State of the Map 2019, 31-32. Heidelberg, Germany, September 21-23, 2019. Available at https://zenodo.org/communities/sotm-2019
DOI: 10.5281/zenodo.338771
Multi-scale satellite images classification : the case of illegal landfills
No full textLAUREA MAGISTRALELa gravità dei problemi connessi alla gestione dei rifiuti a livello mondiale ha favorito in questo settore l’applicazione di metodologie semi-automatiche, almeno a partire dagli anni ‘70. Problema nel problema per i risvolti ambientali e criminali, è rappresentato dalle discariche illegali e in particolare dalla loro individuazione. Per questo obiettivo è stata sfruttata in un primo momento la combinazione di immagini satellitari e sistemi informativi geografici (GIS). Nonostante i buoni risultati ottenuti, soprattutto a partire dal 2000 quando le immagini satellitari sono diventate disponibili in maggior numero e qualità, tali metodi rimangono pur sempre dipendenti dall’intervento umano, con i problemi connessi. Un decisivo miglioramento verso sistemi completamente automatici si è ottenuto con l’adozione di tecniche di deep learning, in particolare le convolutional neural networks (CNNs). In questo ambito le sperimentazioni attuali si focalizzano sulla classificazione automatica delle immagini.
L'obiettivo di questa ricerca è aumentare significativamente le prestazioni di classificazione di una delle reti neurali (ResNet50) già utilizzate in questo ambito. Per ottenere questo risultato, è stata sfruttata la combinazione di informazioni spaziali e semantiche. A tale scopo è stata implementata un'architettura ispirata a Feature Pyramid Network (FPN) che sfrutta l'architettura di base (ResNet50) come backbone. Questa rete permette di superare il problema del rilevamento multiscala, ovvero il riconoscimento di oggetti appartenenti alla stessa classe rappresentati a scale differenti. Il rilevamento multiscala è uno dei principali vincoli all'accuratezza nella classificazione delle immagini tramite CNNs. Gli esperimenti hanno coinvolto 12 diverse configurazioni dell'architettura ResNet50 + FPN, superando alcune criticità legate al rilevamento multiscala. Il modello proposto riesce ad aumentare l'F1 score da 0.865 a 0.882 e l'accuracy di classificazione dal 92% al 93% rispetto all'architettura di base. Lo studio prevede anche una valutazione qualitativa finale con l'obiettivo di identificare sulle immagini le variabili che hanno contribuito a tale risultato.The severity of the worldwide problems associated with waste management has favored the application of semi-automatic methodologies in this sector, at least since the 1970s. Problem within the problem for environmental and criminal implications is represented by illegal landfills and in particular by their identification. For this purpose, the combination of satellite images and geographic information systems (GIS) was first exploited. Despite the good results obtained, especially since 2000 when satellite images became available in greater numbers and quality, these methods still remain dependent on human expertise, with the associated problems. A decisive improvement towards fully automatic systems has been achieved with the adoption of deep learning techniques, in particular Convolutional Neural Networks (CNNs). In this context, current experiments focus on the automatic classification of images.
The goal of this research is to significantly increase the classification performance of one of the neural networks (ResNet50) already used in this field. To achieve this result, the combination of spatial and semantic information was exploited. For this purpose, an architecture inspired by Feature Pyramid Network (FPN) which exploits the baseline architecture (ResNet50) as the backbone, has been implemented. This network allows to overcome the problem of multi-scale detection, that is the recognition of objects belonging to the same class represented at different scales. Multi-scale detection is one of the main constraints on the accuracy in the classification of images via CNNs. The experiments involved 12 different configurations of the ResNet50 + FPN architecture, overcoming some critical issues related to multi-scale detection. The proposed model manages to increase the F1 score from 0.865 to 0.882 and the classification accuracy from 92% to 93% with respect to the baseline architecture. The study also includes a final qualitative evaluation with the aim of identifying on the images the variables that contributed to this result
Attention methods in remote sensing scene classification : the case of illegal landfills
Get PDFLAUREA MAGISTRALELe discariche illegali sono diventate uno dei business più redditizi per le
mafie, e un peso via via crescente per l’economia, l’ambiente e sopratutto
la salute dei cittadini. Al fine di identificarne tempestivamente questi siti
ed evitarne i danni, molte ricerche si sono concentrate sull’automazione del
processo di rilevamento preventivo. Tra questi, uno degli approcci più moderni è l’uso di modelli di apprendimento profondo basati su Reti Neurali
Convoluzionali (CNN), che potrebbero consentire campagne di monitoraggio del territorio su larga scala. In particolare, l’utilizzo dell’architettura
ResNet50 ha già dato buoni risultati.
Questa tesi si propone di valutare gli effetti dell’aggiunta di metodi di
attenzione alla suddetta rete già adoperata per questo task. Questi ultimi
consistono in tecniche per potenziare una CNN, concentrando le risorse computazionali sulle parti più significative dei dati in input. Il loro impiego nel
rilevamento di discariche illegali potrebbe essere utile per un’identificazione
più precisa dei singoli rifiuti presenti nelle immagini. Tra questi meccanismi, lo Squeeze-and-Excitation (SE), il Convolutional Block Attention Module (CBAM) e l’Efficient Channel Attention (ECA) sono stati implementati
sull’architettura esistente di ResNet50. Di questi sono state sperimentate
molte configurazioni, ed è stata eseguita una valutazione quantitativa e qualitativa che ha dimostrato come ECA sia la migliore opzione, in grado di
migliorare le prestazioni di classificazione.
Infine sono state analizzate le Class Activation Maps (CAMs), che sono
heatmaps atte ad evidenziare le aree di un’immagine che determinano maggiormente la classificazione, al fine di comprendere meglio la capacità dei
vari modelli nell’identificare gli oggetti rilevanti nelle scene contenenti discariche abusive. Un dataset ad-hoc è stato creato annotando manualmente
i rifiuti presenti nelle immagini ed un’analisi quantitativa è stata fatta confrontando le CAMs con queste annotazioni. I risultati hanno mostrato che le
CAMs potrebbero essere utili per effettuare rilevamento degli oggetti semisupervisionato, e che ECA è il modello di attenzione migliore.Illegal landfills have become one of the most profitable businesses for criminal organizations and an increasing burden on the economy, the environment
and, above all, the health of citizens. In the fight against this phenomenon,
in order to identify these sites at an early stage and prevent damage, ongoing
research is focusing on automating the process of illegal dumps detection.
Among them, one of the more modern approaches is the use of Deep Learning models based on Convolutional Neural Networks (CNNs), that could
enable mass-scale territory monitoring campaigns. In particular, using the
ResNet50 architecture has already delivered good results.
This thesis evaluates the effects of adding Attention mechanisms to the
above mentioned network. These are techniques designed to enhance a CNN
by focusing its computational resources on the most significant parts of the
input data. Their application on this task of illegal dumps detection could
be useful for a more precise identification of the single wastes present in
the images. In particular, between these mechanisms, the Squeeze-andExcitation (SE), the Convolutional Block Attention module (CBAM) and
the Efficient Channel Attention (ECA) have been implemented on top of
the existing ResNet50 architecture. Several configurations of these models
were tested, and a quantitative and qualitative evaluation showed ECA to
be the best option, obtaining the larger improvement on the classification
performance.
Finally, the Class Activation Maps (CAMs), that are heatmaps designed
to highlight the areas in an image that contributes the most to the classification, have been analyzed, to better understand the various models’
capacity to identify the relevant objects in an illegal landfill scene. An adhoc dataset has been created with annotations of the relevant waste objects,
and a quantitative analysis has been made comparing the CAMs with these
ground truths. The results proved that the CAMs could be the first step
towards weakly-supervised object detection and that, consistently with the
previous results, ECA is the most effective attention module
Suspicious objects classification for illegal landfills discovery in remote sensing images
No full textLAUREA MAGISTRALEIl problema della gestione dei rifiuti ha recentemente acquisito rilevanza a livello mondiale, con una risonanza sia economica che sociale in ogni Paese. Una delle criticità più preoccupanti è costituita dalle discariche abusive, ovvero nello scarico incontrollato di rifiuti nell'ambiente. Dagli anni '70, con l'ulteriore sviluppo della tecnologia, la necessità di monitorare in maniera automatica questi tipi di fenomeni è aumentata, portando i ricercatori a esplorare molte possibili opzioni. Finora, però, non è stato possibile andare oltre a tecniche semiautomatiche, che richiedono sempre l'intervento umano. La combinazione delle immagini satellitari con i sistemi di informazione geografica (GIS) è stata esaminata in lungo e in largo per molto tempo. Nonostante abbia prodotto ottimi risultati, queste tecniche non hanno mai permesso di raggiungere la totale indipendenza dalle competenze umane. Un deciso miglioramento verso i sistemi completamente automatici è stato ottenuto con l'adozione di tecniche di Deep Learning, in particolare le Convolutional Neural Networks (CNN). In questo contesto, gli esperimenti attuali si concentrano sulla classificazione automatica delle immagini.
L'obiettivo di questa ricerca è sfruttare una delle reti neurali già utilizzate in questo campo (ResNet50) per estendere ulteriormente le capacità di monitoraggio di sistemi automatici, consentendo la classificazione di diverse tipologie di oggetti che caratterizzano le discariche illegali. Inizialmente l'architettura ResNet50 è stata utilizzata per risolvere la classificazione di tipo multilabel, e quindi per riconoscere la presenza (anche contemporanea) di oggetti appartenenti a diverse classi. Il modello proposto ha raggiunto un F1 score del 81% in media sul test set. Successivamente, con il classificatore allenato sono state prodotte le Class Activation Maps (CAMs), al fine di identificare le regioni delle immagini che appartengono ai diversi tipi di rifiuti. I risultati ottenuti sono stati valutati quantitativamente utilizzando metriche personalizzate basate sul calcolo dell'Intersection over Union, e qualitativamente guardando effettivamente i box di delimitazione ottenuti.The problem of waste management has recently gained worldwide relevance, having both an economic and social resonance in every country. One of the most concerning issues is constituted by illegal dumping, consisting of the uncontrolled discharge of waste into the environment. From the 1970s, with the further development of technology, the need for automatic procedures to monitor these types of phenomena increased over and over, bringing researchers to explore many possible options. Until now, however, it has been not possible to go beyond semi-automatic techniques, always requiring human intervention. In particular, the combination of satellite images with geographic information systems (GIS) has been examined far and wide for a long time. Even if it produced very good results, as anticipated, it has never allowed achieving total independence from human expertise. A decisive improvement towards fully automatic systems has been obtained with the adoption of Deep Learning techniques, in particular Convolutional Neural Networks (CNNs). In this context, current experiments focus on the automatic classification of images.
The goal of this research is to exploit one of the neural networks already used in this field, the ResNet50 architecture, to further extend the monitoring capabilities of automatic systems, allowing them to also account for the classification and localization of different types of objects characterizing Illegal Landfills. In particular, firstly the ResNet50 architecture has been used to solve the multilabel classification task, consisting in the recognition of the presence (even concurrently) of the considered classes in the given images. The proposed model reached an F1 score of 81% on average on the test set. From the trained classifier, Class Activation Maps (CAMs) were produced and analyzed to identify the regions of the images belonging to the different waste types. The obtained results have been evaluated quantitatively using custom metrics based on the calculation of the Intersection over Union, and qualitatively by actually looking at the obtained bounding boxes, thus understanding the practical relevance of the achieved results
Learning to find mountains
Get PDFLAUREA MAGISTRALEAnalizzare dati digitali, per identificare e classificare la morfologia del terreno, è un compito importante, che può contribuire a migliorare la disponibilità e la qualità della cartografia pubblica open source e a sviluppare nuove applicazioni per il monitoraggio del turismo e dell'ambiente. Nella letteratura, sono documentati alcuni algoritmi euristici per identificare caratteristiche di regioni montane, soprattutto le coordinate delle vette, grazie a set di dati in input, come il Modello Digitale di Elevazione (DEM) della Terra. Scegliere il metodo, da utilizzare per l'individuazione dei picchi delle montagne, dipende dai requisiti della applicazione a portata di mano; ma la decisione è aiutata anche dalla disponibilità di un rigoroso confronto tra i differenti metodi. Tutti questi algoritmi dipendono da parametri che sono da impostare manualmente. In questa tesi, esploriamo l'uso di metodi di Deep Learning, in particolare classificazione e segmentazione, per addestrare un modello, in grado di identificare vette di montagne, e che impari da un set di dati, usato come gold standard, contenente le coordinate dei picchi in una regione. I modelli sono stati addestrati e provati con dati DEM e picchi della Svizzera. Inoltre, proponiamo un approccio per un confronto equo, in termini di Precision-Recall e errore medio sulla distanza, e presentiamo i risultati quantitativi e qualitativi ottenuti valutando, in una area della Svizzera, i metodi più noti dalla letteratura e i nostri modelli di Deep Learning.Analyzing digital data to identify and classify landforms is an important task, which can contribute to improve the availability and quality of public open source cartography and to develop novel applications for tourism and environment monitoring. In the literature, several heuristic algorithms are documented for identifying the features of mountain regions, most notably the coordinates of summits, from input datasets, such as the Digital Elevation Model (DEM) of the Earth. Choosing the method to use for mountain peaks detection depends on the requirements of the application at hand, but the decision is helped also by the availability of a rigorous comparison among the different methods. All these algorithms depend on parameters, which are manually set. In this thesis, we explore the use of Deep Learning methods, specifically classification and segmentation, to train a model capable of identifying mountain summits, which learns from a gold standard dataset containing the coordinates of peaks in a region. The models have been trained and tested with Switzerland DEM and peak data. Furthermore, we propose an approach for a fair comparison, in terms of Precision-Recall and mean distance error, and present the quantitative and qualitative results obtained evaluating the best known methods from the literature and our Deep Learning models, in an area of Switzerland
- …
