Continual learning for computer vision applications

One of the most visionary goals of Artificial Intelligence is to create a system able to mimic and eventually surpass the intelligence observed in biological systems including, ambitiously, the one observed in humans. The main distinctive strength of humans is their ability to build a deep understanding of the world by learning continuously and drawing from their experiences. This ability, which is found in various degrees in all intelligent biological beings, allows them to adapt and properly react to changes by incrementally expanding and refining their knowledge. Arguably, achieving this ability is one of the main goals of Artificial Intelligence and a cornerstone towards the creation of intelligent artificial agents. Modern Deep Learning approaches allowed researchers and industries to achieve great advancements towards the resolution of many long-standing problems in areas like Computer Vision and Natural Language Processing. However, while this current age of renewed interest in AI allowed for the creation of extremely useful applications, a concerningly limited effort is being directed towards the design of systems able to learn continuously. The biggest problem that hinders an AI system from learning incrementally is the catastrophic forgetting phenomenon. This phenomenon, which was discovered in the 90s, naturally occurs in Deep Learning architectures where classic learning paradigms are applied when learning incrementally from a stream of experiences. This dissertation revolves around the Continual Learning field, a sub-field of Machine Learning research that has recently made a comeback following the renewed interest in Deep Learning approaches. This work will focus on a comprehensive view of continual learning by considering algorithmic, benchmarking, and applicative aspects of this field. This dissertation will also touch on community aspects such as the design and creation of research tools aimed at supporting Continual Learning research, and the theoretical and practical aspects concerning public competitions in this field.

Computer vision in aseptic pharmaceutical manufacturing: a deep learning-based approach to vial inspection and pack formation during a lyophilizer loading cycle

Vision systems are powerful tools playing an increasingly important role in modern industry, to detect errors and maintain product standards. With the enlarged availability of affordable industrial cameras, computer vision algorithms have been increasingly applied in industrial manufacturing processes monitoring. Until a few years ago, industrial computer vision applications relied only on ad-hoc algorithms designed for the specific object and acquisition setup being monitored, with a strong focus on co-designing the acquisition and processing pipeline. Deep learning has overcome these limits providing greater flexibility and faster re-configuration. In this work, the process to be inspected consists in vials’ pack formation entering a freeze-dryer, which is a common scenario in pharmaceutical active ingredient packaging lines. To ensure that the machine produces proper packs, a vision system is installed at the entrance of the freeze-dryer to detect eventual anomalies with execution times compatible with the production specifications. Other constraints come from sterility and safety standards required in pharmaceutical manufacturing. This work presents an overview about the production line, with particular focus on the vision system designed, and about all trials conducted to obtain the final performance. Transfer learning, alleviating the requirement for a large number of training data, combined with data augmentation methods, consisting in the generation of synthetic images, were used to effectively increase the performances while reducing the cost of data acquisition and annotation. The proposed vision algorithm is composed by two main subtasks, designed respectively to vials counting and discrepancy detection. The first one was trained on more than 23k vials (about 300 images) and tested on 5k more (about 75 images), whereas 60 training images and 52 testing images were used for the second one.

Geophotohunt: Studio di algoritmi di confronto di immagini per realizzare una caccia al tesoro fotografica

Questo studio si propone di realizzare un’applicazione per dispositivi Android che permetta, per mezzo di un gioco di ruolo strutturato come caccia al tesoro, di visitare in prima persona città d’arte e luoghi turistici. Gli utenti finali, grazie alle funzionalità dell’app stessa, potranno giocare, creare e condividere cacce al tesoro basate sulla ricerca di edifici, monumenti, luoghi di rilevanza artistico-storica o turistica; in particolare al fine di completare ciascuna tappa di una caccia al tesoro il giocatore dovrà scattare una fotografia al monumento o edificio descritto nell’obiettivo della caccia stessa. Il software grazie ai dati rilevati tramite GPS e giroscopio (qualora il dispositivo ne sia dotato) e per mezzo di un algoritmo di instance recognition sarà in grado di affermare se la foto scattata rappresenta la risposta corretta al quesito della tappa. L’applicazione GeoPhotoHunt rappresenta non solo uno strumento ludico per la visita di città turistiche o più in generale luoghi di interesse, lo studio propone, infatti come suo contributo originale, l’implementazione su piattaforma mobile di un Content Based Image Retrieval System (CBIR) del tutto indipendente da un supporto server. Nello specifico il server dell’applicazione non sarà altro che uno strumento di appoggio con il quale i membri della “community” di GeoPhotoHunt potranno pubblicare le cacce al tesoro da loro create e condividere i punteggi che hanno totalizzato partecipando a una caccia al tesoro. In questo modo quando un utente ha scaricato sul proprio smartphone i dati di una caccia al tesoro potrà iniziare l’avventura anche in assenza di una connessione internet. L’intero studio è stato suddiviso in più fasi, ognuna di queste corrisponde ad una specifica sezione dell’elaborato che segue. In primo luogo si sono effettuate delle ricerche, soprattutto nel web, con lo scopo di individuare altre applicazioni che implementano l’idea della caccia al tesoro su piattaforma mobile o applicazioni che implementassero algoritmi di instance recognition direttamente su smartphone. In secondo luogo si è ricercato in letteratura quali fossero gli algoritmi di riconoscimento di immagini più largamente diffusi e studiati in modo da avere una panoramica dei metodi da testare per poi fare la scelta dell’algoritmo più adatto al caso di studio. Quindi si è proceduto con lo sviluppo dell’applicazione GeoPhotoHunt stessa, sia per quanto riguarda l’app front-end per dispositivi Android sia la parte back-end server. Infine si è passati ad una fase di test di algoritmi di riconoscimento di immagini in modo di avere una sufficiente quantità di dati sperimentali da permettere di effettuare una scelta dell’algoritmo più adatto al caso di studio. Al termine della fase di testing si è deciso di implementare su Android un algoritmo basato sulla distanza tra istogrammi di colore costruiti sulla scala cromatica HSV, questo metodo pur non essendo robusto in presenza di variazioni di luminosità e contrasto, rappresenta un buon compromesso tra prestazioni, complessità computazionale in modo da rendere la user experience quanto più coinvolgente.

Geophotohunt: Progettazione e sviluppo di una App Android per la gestione di cacce al tesoro fotografiche

Il problema che si vuole affrontare è la progettazione e lo sviluppo di un sistema interattivo volto all’apprendimento e alla visita guidata di città d’arte. Si vuole realizzare un’applicazione per dispositivi mobili che offra sia il servizio di creazione di visite guidate che l’utilizzo delle stesse in assenza di connessione internet. Per rendere l’utilizzo dei servizi offerti più piacevole e divertente si è deciso di realizzare le visite guidate sotto forma di cacce al tesoro fotografiche, le cui tappe consistono in indizi testuali che per essere risolti richiedono risposte di tipo fotografico. Si è inoltre scelto di realizzare una community volta alla condivisione delle cacce al tesoro realizzate e al mantenimento di statistiche di gioco. Il contributo originale di questa tesi consiste nella progettazione e realizzazione di una App Android, denominata GeoPhotoHunt, che sfrutta l’idea della caccia al tesoro fotografica e geo localizzata per facilitare le visite guidate a luoghi di interesse, senza la necessità di una connessione ad internet. Il client viene reso indipendente dal server grazie allo spostamento degli algoritmi di image recognition sul client. Esentare il client dalla necessità di una connessione ad internet permette il suo utilizzo anche in città estere dove solitamente non si ha possibilità di connettersi alla rete.

A Real Time Vision System for Autonomous Systems: Characterization during a Middle Size Match

in RoboCup 2007: Robot Soccer World Cup XI

Vision-based hand Wheel-chair control

Several studies have shown that people with disabilities benefit substantially from access to a means of independent mobility and assistive technology. Researchers are using technology originally developed for mobile robots to create easier to use wheelchairs. With this kind of technology people with disabilities can gain a degree of independence in performing daily life activities. In this work a computer vision system is presented, able to drive a wheelchair with a minimum number of finger commands. The user hand is detected and segmented with the use of a kinect camera, and fingertips are extracted from depth information, and used as wheelchair commands.

Hand gesture recognition for human computer interaction : a comparative study of different image features

Hand gesture recognition for human computer interaction, being a natural way of human computer interaction, is an area of active research in computer vision and machine learning. This is an area with many different possible applications, giving users a simpler and more natural way to communicate with robots/systems interfaces, without the need for extra devices. So, the primary goal of gesture recognition research is to create systems, which can identify specific human gestures and use them to convey information or for device control. For that, vision-based hand gesture interfaces require fast and extremely robust hand detection, and gesture recognition in real time. In this study we try to identify hand features that, isolated, respond better in various situations in human-computer interaction. The extracted features are used to train a set of classifiers with the help of RapidMiner in order to find the best learner. A dataset with our own gesture vocabulary consisted of 10 gestures, recorded from 20 users was created for later processing. Experimental results show that the radial signature and the centroid distance are the features that when used separately obtain better results, with an accuracy of 91% and 90,1% respectively obtained with a Neural Network classifier. These to methods have also the advantage of being simple in terms of computational complexity, which make them good candidates for real-time hand gesture recognition.

Generic system for human-computer gesture interaction: applications on sign language recognition and robotic soccer refereeing

Hand gestures are a powerful way for human communication, with lots of potential applications in the area of human computer interaction. Vision-based hand gesture recognition techniques have many proven advantages compared with traditional devices, giving users a simpler and more natural way to communicate with electronic devices. This work proposes a generic system architecture based in computer vision and machine learning, able to be used with any interface for human-computer interaction. The proposed solution is mainly composed of three modules: a pre-processing and hand segmentation module, a static gesture interface module and a dynamic gesture interface module. The experiments showed that the core of visionbased interaction systems could be the same for all applications and thus facilitate the implementation. For hand posture recognition, a SVM (Support Vector Machine) model was trained and used, able to achieve a final accuracy of 99.4%. For dynamic gestures, an HMM (Hidden Markov Model) model was trained for each gesture that the system could recognize with a final average accuracy of 93.7%. The proposed solution as the advantage of being generic enough with the trained models able to work in real-time, allowing its application in a wide range of human-machine applications. To validate the proposed framework two applications were implemented. The first one is a real-time system able to interpret the Portuguese Sign Language. The second one is an online system able to help a robotic soccer game referee judge a game in real time.

Vision-based portuguese sign language recognition system

Vision-based hand gesture recognition is an area of active current research in computer vision and machine learning. Being a natural way of human interaction, it is an area where many researchers are working on, with the goal of making human computer interaction (HCI) easier and natural, without the need for any extra devices. So, the primary goal of gesture recognition research is to create systems, which can identify specific human gestures and use them, for example, to convey information. For that, vision-based hand gesture interfaces require fast and extremely robust hand detection, and gesture recognition in real time. Hand gestures are a powerful human communication modality with lots of potential applications and in this context we have sign language recognition, the communication method of deaf people. Sign lan- guages are not standard and universal and the grammars differ from country to coun- try. In this paper, a real-time system able to interpret the Portuguese Sign Language is presented and described. Experiments showed that the system was able to reliably recognize the vowels in real-time, with an accuracy of 99.4% with one dataset of fea- tures and an accuracy of 99.6% with a second dataset of features. Although the im- plemented solution was only trained to recognize the vowels, it is easily extended to recognize the rest of the alphabet, being a solid foundation for the development of any vision-based sign language recognition user interface system.

Generic system for human-computer gesture interaction

Hand gestures are a powerful way for human communication, with lots of potential applications in the area of human computer interaction. Vision-based hand gesture recognition techniques have many proven advantages compared with traditional devices, giving users a simpler and more natural way to communicate with electronic devices. This work proposes a generic system architecture based in computer vision and machine learning, able to be used with any interface for humancomputer interaction. The proposed solution is mainly composed of three modules: a pre-processing and hand segmentation module, a static gesture interface module and a dynamic gesture interface module. The experiments showed that the core of vision-based interaction systems can be the same for all applications and thus facilitate the implementation. In order to test the proposed solutions, three prototypes were implemented. For hand posture recognition, a SVM model was trained and used, able to achieve a final accuracy of 99.4%. For dynamic gestures, an HMM model was trained for each gesture that the system could recognize with a final average accuracy of 93.7%. The proposed solution as the advantage of being generic enough with the trained models able to work in real-time, allowing its application in a wide range of human-machine applications.

AppRecommender : disseny i desenvolupament d'una aplicació per a Android

Aquest document és la memòria que recull el procés de disseny i desenvolupament d'una aplicació per a dispositius Android, que he realitzat com a projecte de fi de carrera de la titulació d'Enginyeria Informàtica.

Accuracy estimation of a new omnidirectional 3D vision sensor

We present a computer vision system that associates omnidirectional vision with structured light with the aim of obtaining depth information for a 360 degrees field of view. The approach proposed in this article combines an omnidirectional camera with a panoramic laser projector. The article shows how the sensor is modelled and its accuracy is proved by means of experimental results. The proposed sensor provides useful information for robot navigation applications, pipe inspection, 3D scene modelling etc

Behaviours implemented using a vision system based on coded light projection

In a search for new sensor systems and new methods for underwater vehicle positioning based on visual observation, this paper presents a computer vision system based on coded light projection. 3D information is taken from an underwater scene. This information is used to test obstacle avoidance behaviour. In addition, the main ideas for achieving stabilisation of the vehicle in front of an object are presented

Desenvolupament d'una aplicació mòbil tipus gimcana per plataformes Android

L'objectiu del projecte és el desenvolupament d'una aplicacióper a mòbils amb plataforma Android que generi una gimcana, per poder realitzar unagimcana més completa i propera a la nostra era actual; incorporant totes lesfuncionalitats que tenen els telèfons mòbils a les eines per resoldre un joc de proves, icrear així un concepte més actual del que és una gimcana tradicional. La creació del'aplicació pretén fomentar i establir un diàleg de saber col•laboratiu i d'aportació deconeixements entre els components del mateix grup, que potencia la implicació en lacreació de la resposta al misteri o la realització de la prova

Learning Grasp Affordances from Vision

Monimutkaisissa ja muuttuvissa ympäristöissä työskentelevät robotit tarvitsevat kykyä manipuloida ja tarttua esineisiin. Tämä työ tutkii robottitarttumisen ja robottitartuntapis-teiden koneoppimisen aiempaa tutkimusta ja nykytilaa. Nykyaikaiset menetelmät käydään läpi, ja Le:n koneoppimiseen pohjautuva luokitin toteutetaan, koska se tarjoaa parhaan onnistumisprosentin tutkituista menetelmistä ja on muokattavissa sopivaksi käytettävissä olevalle robotille. Toteutettu menetelmä käyttää intensititeettikuvaan ja syvyyskuvaan po-hjautuvia ominaisuuksi luokitellakseen potentiaaliset tartuntapisteet. Tämän toteutuksen tulokset esitellään.