Breaking Waves on the Ocean Surface

Thesis (Ph.D.)--University of Washington, 2016-08

Problems in Algebraic Vision

Thesis (Ph.D.)--University of Washington, 2016-08

Panoramic Sea Happening (After Kantor): Video Installation

Panoramic Sea Happening (After Kantor) is a 7 minute durational film that reimagines part of Tadeusz Kantor's original sea happenings from 1967 in a landscape in which the sea has retreated. The conductor of Kantor’s original performance is replaced with a sound object cast adrift on a beach in Dungeness (UK). The object plays back the sound of the sea into the landscape, which was performed live and then filmed from three distinct angles. The first angle mimics the position of the conductor in Kantor’s original happening, facing outwards into the horizon of the beach and recalls the image in Kantor’s work of a human figure undertaking the absurd task of orchestrating the sound of a gigantic expanse of water. The second angle exposes the machine itself and the large cone that amplifies the sound, reinforcing the isolation of the object. The third angle reveals a decommissioned nuclear power station and sound objects used as a warning system for the power plant. Dungeness is a location where the sea has been retreating from the land, leaving traces of human activity through the disused boat winches, abandoned cabins and the decommissioned nuclear buildings. It is a place in which the footprint of the anthropocene is keenly felt. The sound object is intended to act as an anthropomorphic figure, ghosting the original conductor and offering the sound of the sea back into the landscape through a wide mouthpiece, echoing Kantor’s own load hailer in the original sequence of sea happenings. It speculates on Kantor's theory of the bio-object, which proposed a symbiotic relationship between the human and the nonhuman object in performance, as a possible instrument to access a form of geologic imagination. In this configuration, the human itself is absent, but is evoked through the objects left behind. The sound object, helpless in a red dingy, might be thought of as a co-conspirator with the viewer, enabling a looking back to the past in a landscape of an inevitable future. The work was originally commissioned by the University of Kent in collaboration with the Polish Cultural Institute for the Symposium Kantorbury Kantorbury in Canterbury (UK) to mark the 100 years since Tadeusz Kantor’s birth (15 - 19 September 2015). It should be projected and requires stereo speakers.

Vision-based tracking for sports performance analysis

With the world of professional sports shifting towards employing better sport analytics, the demand for vision-based performance analysis is growing increasingly in recent years. In addition, the nature of many sports does not allow the use of any kind of sensors or other wearable markers attached to players for monitoring their performances during competitions. This provides a potential application of systematic observations such as tracking information of the players to help coaches to develop their visual skills and perceptual awareness needed to make decisions about team strategy or training plans. My PhD project is part of a bigger ongoing project between sport scientists and computer scientists involving also industry partners and sports organisations. The overall idea is to investigate the contribution technology can make to the analysis of sports performance on the example of team sports such as rugby, football or hockey. A particular focus is on vision-based tracking, so that information about the location and dynamics of the players can be gained without any additional sensors on the players. To start with, prior approaches on visual tracking are extensively reviewed and analysed. In this thesis, methods to deal with the difficulties in visual tracking to handle the target appearance changes caused by intrinsic (e.g. pose variation) and extrinsic factors, such as occlusion, are proposed. This analysis highlights the importance of the proposed visual tracking algorithms, which reflect these challenges and suggest robust and accurate frameworks to estimate the target state in a complex tracking scenario such as a sports scene, thereby facilitating the tracking process. Next, a framework for continuously tracking multiple targets is proposed. Compared to single target tracking, multi-target tracking such as tracking the players on a sports field, poses additional difficulties, namely data association, which needs to be addressed. Here, the aim is to locate all targets of interest, inferring their trajectories and deciding which observation corresponds to which target trajectory is. In this thesis, an efficient framework is proposed to handle this particular problem, especially in sport scenes, where the players of the same team tend to look similar and exhibit complex interactions and unpredictable movements resulting in matching ambiguity between the players. The presented approach is also evaluated on different sports datasets and shows promising results. Finally, information from the proposed tracking system is utilised as the basic input for further higher level performance analysis such as tactics and team formations, which can help coaches to design a better training plan. Due to the continuous nature of many team sports (e.g. soccer, hockey), it is not straightforward to infer the high-level team behaviours, such as players’ interaction. The proposed framework relies on two distinct levels of performance analysis: low-level performance analysis, such as identifying players positions on the play field, as well as a high-level analysis, where the aim is to estimate the density of player locations or detecting their possible interaction group. The related experiments show the proposed approach can effectively explore this high-level information, which has many potential applications.

Datasets, features, learning, and models in visual recognition

Visual recognition is a fundamental research topic in computer vision. This dissertation explores datasets, features, learning, and models used for visual recognition. In order to train visual models and evaluate different recognition algorithms, this dissertation develops an approach to collect object image datasets on web pages using an analysis of text around the image and of image appearance. This method exploits established online knowledge resources (Wikipedia pages for text; Flickr and Caltech data sets for images). The resources provide rich text and object appearance information. This dissertation describes results on two datasets. The first is Berg’s collection of 10 animal categories; on this dataset, we significantly outperform previous approaches. On an additional set of 5 categories, experimental results show the effectiveness of the method. Images are represented as features for visual recognition. This dissertation introduces a text-based image feature and demonstrates that it consistently improves performance on hard object classification problems. The feature is built using an auxiliary dataset of images annotated with tags, downloaded from the Internet. Image tags are noisy. The method obtains the text features of an unannotated image from the tags of its k-nearest neighbors in this auxiliary collection. A visual classifier presented with an object viewed under novel circumstances (say, a new viewing direction) must rely on its visual examples. This text feature may not change, because the auxiliary dataset likely contains a similar picture. While the tags associated with images are noisy, they are more stable when appearance changes. The performance of this feature is tested using PASCAL VOC 2006 and 2007 datasets. This feature performs well; it consistently improves the performance of visual object classifiers, and is particularly effective when the training dataset is small. With more and more collected training data, computational cost becomes a bottleneck, especially when training sophisticated classifiers such as kernelized SVM. This dissertation proposes a fast training algorithm called Stochastic Intersection Kernel Machine (SIKMA). This proposed training method will be useful for many vision problems, as it can produce a kernel classifier that is more accurate than a linear classifier, and can be trained on tens of thousands of examples in two minutes. It processes training examples one by one in a sequence, so memory cost is no longer the bottleneck to process large scale datasets. This dissertation applies this approach to train classifiers of Flickr groups with many group training examples. The resulting Flickr group prediction scores can be used to measure image similarity between two images. Experimental results on the Corel dataset and a PASCAL VOC dataset show the learned Flickr features perform better on image matching, retrieval, and classification than conventional visual features. Visual models are usually trained to best separate positive and negative training examples. However, when recognizing a large number of object categories, there may not be enough training examples for most objects, due to the intrinsic long-tailed distribution of objects in the real world. This dissertation proposes an approach to use comparative object similarity. The key insight is that, given a set of object categories which are similar and a set of categories which are dissimilar, a good object model should respond more strongly to examples from similar categories than to examples from dissimilar categories. This dissertation develops a regularized kernel machine algorithm to use this category dependent similarity regularization. Experiments on hundreds of categories show that our method can make significant improvement for categories with few or even no positive examples.

Binocular functional architecture for detection of contrast-modulated gratings

Combination of signals from the two eyes is the gateway to stereo vision. To gain insight into binocular signal processing, we studied binocular summation for luminance-modulated gratings (L or LM) and contrast-modulated gratings (CM). We measured 2AFC detection thresholds for a signal grating (0.75 c/deg, 216msec) shown to one eye, both eyes, or both eyes out-of-phase. For LM and CM, the carrier noise was in both eyes, even when the signal was monocular. Mean binocular thresholds for luminance gratings (L) were 5.4dB better than monocular thresholds - close to perfect linear summation (6dB). For LM and CM the binocular advantage was again 5-6dB, even when the carrier noise was uncorrelated, anti-correlated, or at orthogonal orientations in the two eyes. Binocular combination for CM probably arises from summation of envelope responses, and not from summation of these conflicting carrier patterns. Antiphase signals produced no binocular advantage, but thresholds were about 1-3dB higher than monocular ones. This is not consistent with simple linear summation, which should give complete cancellation and unmeasurably high thresholds. We propose a three-channel model in which noisy monocular responses to the envelope are binocularly combined in a contrast-weighted sum, but also remain separately available to perception via a max operator. Vision selects the largest of the three responses. With in-phase gratings the binocular channel dominates, but antiphase gratings cancel in the binocular channel and the monocular channels mediate detection. The small antiphase disadvantage might be explained by a subtle influence of background responses on binocular and monocular detection.

Contrast and lustre:a model that accounts for eleven different forms of contrast discrimination in binocular vision

Our goal here is a more complete understanding of how information about luminance contrast is encoded and used by the binocular visual system. In two-interval forced-choice experiments we assessed observers' ability to discriminate changes in contrast that could be an increase or decrease of contrast in one or both eyes, or an increase in one eye coupled with a decrease in the other (termed IncDec). The base or pedestal contrasts were either in-phase or out-of-phase in the two eyes. The opposed changes in the IncDec condition did not cancel each other out, implying that along with binocular summation, information is also available from mechanisms that do not sum the two eyes' inputs. These might be monocular mechanisms. With a binocular pedestal, monocular increments of contrast were much easier to see than monocular decrements. These findings suggest that there are separate binocular (B) and monocular (L,R) channels, but only the largest of the three responses, max(L,B,R), is available to perception and decision. Results from contrast discrimination and contrast matching tasks were described very accurately by this model. Stimuli, data, and model responses can all be visualized in a common binocular contrast space, allowing a more direct comparison between models and data. Some results with out-of-phase pedestals were not accounted for by the max model of contrast coding, but were well explained by an extended model in which gratings of opposite polarity create the sensation of lustre. Observers can discriminate changes in lustre alongside changes in contrast.

Visual suppression in intermittent exotropia during binocular alignment.

PURPOSE To investigate the cortical mechanisms that prevent diplopia in intermittent exotropia (X(T)) during binocular alignment (orthotropia). METHODS The authors studied 12 X(T) patients aged 5 to 22 years. Seventy-five percent had functional stereo vision with stereoacuity similar to that of 12 age-matched controls (0.2-3.7 min arc). Identical face images were presented to the two eyes for 400 ms. In one eye, the face was presented at the fovea; in the other, offset along the horizontal axis with up to 12° eccentricity. The task was to indicate whether one or two faces were perceived. RESULTS All X(T) patients showed normal diplopia when the nonfoveal face was presented to nasal hemiretina, though with a slightly larger fusional range than age-matched controls. However, 10 of 12 patients never experienced diplopia when the nonfoveal face was presented to temporal hemiretina (i.e., when the stimulus simulated exodeviation). Patients showed considerable variability when the single image was perceived. Some patients suppressed the temporal stimulus regardless of which eye viewed it, whereas others suppressed a particular eye even when it viewed the foveal stimulus. In two patients, the simulated exodeviation might have triggered a shift from normal to anomalous retinal correspondence. CONCLUSIONS Antidiplopic mechanisms in X(T) can be reliably triggered by purely retinal information during orthotropia, but the nature of these mechanisms varies between patients.

Optimization of Pattern Matching Algorithms for Multi- and Many-Core Platforms

Image and video compression play a major role in the world today, allowing the storage and transmission of large multimedia content volumes. However, the processing of this information requires high computational resources, hence the improvement of the computational performance of these compression algorithms is very important. The Multidimensional Multiscale Parser (MMP) is a pattern-matching-based compression algorithm for multimedia contents, namely images, achieving high compression ratios, maintaining good image quality, Rodrigues et al. [2008]. However, in comparison with other existing algorithms, this algorithm takes some time to execute. Therefore, two parallel implementations for GPUs were proposed by Ribeiro [2016] and Silva [2015] in CUDA and OpenCL-GPU, respectively. In this dissertation, to complement the referred work, we propose two parallel versions that run the MMP algorithm in CPU: one resorting to OpenMP and another that converts the existing OpenCL-GPU into OpenCL-CPU. The proposed solutions are able to improve the computational performance of MMP by 3 and 2:7 , respectively. The High Efficiency Video Coding (HEVC/H.265) is the most recent standard for compression of image and video. Its impressive compression performance, makes it a target for many adaptations, particularly for holoscopic image/video processing (or light field). Some of the proposed modifications to encode this new multimedia content are based on geometry-based disparity compensations (SS), developed by Conti et al. [2014], and a Geometric Transformations (GT) module, proposed by Monteiro et al. [2015]. These compression algorithms for holoscopic images based on HEVC present an implementation of specific search for similar micro-images that is more efficient than the one performed by HEVC, but its implementation is considerably slower than HEVC. In order to enable better execution times, we choose to use the OpenCL API as the GPU enabling language in order to increase the module performance. With its most costly setting, we are able to reduce the GT module execution time from 6.9 days to less then 4 hours, effectively attaining a speedup of 45 .

Scene Mapping and Understanding by Robotic Vision

The first mechanical Automaton concept was found in a Chinese text written in the 3rd century BC, while Computer Vision was born in the late 1960s. Therefore, visual perception applied to machines (i.e. the Machine Vision) is a young and exciting alliance. When robots came in, the new field of Robotic Vision was born, and these terms began to be erroneously interchanged. In short, we can say that Machine Vision is an engineering domain, which concern the industrial use of Vision. The Robotic Vision, instead, is a research field that tries to incorporate robotics aspects in computer vision algorithms. Visual Servoing, for example, is one of the problems that cannot be solved by computer vision only. Accordingly, a large part of this work deals with boosting popular Computer Vision techniques by exploiting robotics: e.g. the use of kinematics to localize a vision sensor, mounted as the robot end-effector. The remainder of this work is dedicated to the counterparty, i.e. the use of computer vision to solve real robotic problems like grasping objects or navigate avoiding obstacles. Will be presented a brief survey about mapping data structures most widely used in robotics along with SkiMap, a novel sparse data structure created both for robotic mapping and as a general purpose 3D spatial index. Thus, several approaches to implement Object Detection and Manipulation, by exploiting the aforementioned mapping strategies, will be proposed, along with a completely new Machine Teaching facility in order to simply the training procedure of modern Deep Learning networks.

Previsione del Successo di Prodotti di Moda Prima della Commercializzazione: un Nuovo Dataset e Modello di Vision-Language Transformer

A differenza di quanto avviene nel commercio tradizionale, in quello online il cliente non ha la possibilità di toccare con mano o provare il prodotto. La decisione di acquisto viene maturata in base ai dati messi a disposizione dal venditore attraverso titolo, descrizioni, immagini e alle recensioni di clienti precedenti. É quindi possibile prevedere quanto un prodotto venderà sulla base di queste informazioni. La maggior parte delle soluzioni attualmente presenti in letteratura effettua previsioni basandosi sulle recensioni, oppure analizzando il linguaggio usato nelle descrizioni per capire come questo influenzi le vendite. Le recensioni, tuttavia, non sono informazioni note ai venditori prima della commercializzazione del prodotto; usando solo dati testuali, inoltre, si tralascia l’influenza delle immagini. L'obiettivo di questa tesi è usare modelli di machine learning per prevedere il successo di vendita di un prodotto a partire dalle informazioni disponibili al venditore prima della commercializzazione. Si fa questo introducendo un modello cross-modale basato su Vision-Language Transformer in grado di effettuare classificazione. Un modello di questo tipo può aiutare i venditori a massimizzare il successo di vendita dei prodotti. A causa della mancanza, in letteratura, di dataset contenenti informazioni relative a prodotti venduti online che includono l’indicazione del successo di vendita, il lavoro svolto comprende la realizzazione di un dataset adatto a testare la soluzione sviluppata. Il dataset contiene un elenco di 78300 prodotti di Moda venduti su Amazon, per ognuno dei quali vengono riportate le principali informazioni messe a disposizione dal venditore e una misura di successo sul mercato. Questa viene ricavata a partire dal gradimento espresso dagli acquirenti e dal posizionamento del prodotto in una graduatoria basata sul numero di esemplari venduti.

End-to-end Deep Metric Learning con Vision-Language Model per il Fashion Image Captioning

L'image captioning è un task di machine learning che consiste nella generazione di una didascalia, o caption, che descriva le caratteristiche di un'immagine data in input. Questo può essere applicato, ad esempio, per descrivere in dettaglio i prodotti in vendita su un sito di e-commerce, migliorando l'accessibilità del sito web e permettendo un acquisto più consapevole ai clienti con difficoltà visive. La generazione di descrizioni accurate per gli articoli di moda online è importante non solo per migliorare le esperienze di acquisto dei clienti, ma anche per aumentare le vendite online. Oltre alla necessità di presentare correttamente gli attributi degli articoli, infatti, descrivere i propri prodotti con il giusto linguaggio può contribuire a catturare l'attenzione dei clienti. In questa tesi, ci poniamo l'obiettivo di sviluppare un sistema in grado di generare una caption che descriva in modo dettagliato l'immagine di un prodotto dell'industria della moda dato in input, sia esso un capo di vestiario o un qualche tipo di accessorio. A questo proposito, negli ultimi anni molti studi hanno proposto soluzioni basate su reti convoluzionali e LSTM. In questo progetto proponiamo invece un'architettura encoder-decoder, che utilizza il modello Vision Transformer per la codifica delle immagini e GPT-2 per la generazione dei testi. Studiamo inoltre come tecniche di deep metric learning applicate in end-to-end durante l'addestramento influenzino le metriche e la qualità delle caption generate dal nostro modello.

Supporting Scientific Research Through Machine and Deep Learning: Fluorescence Microscopy and Operational Intelligence Use Cases

Although the debate of what data science is has a long history and has not reached a complete consensus yet, Data Science can be summarized as the process of learning from data. Guided by the above vision, this thesis presents two independent data science projects developed in the scope of multidisciplinary applied research. The first part analyzes fluorescence microscopy images typically produced in life science experiments, where the objective is to count how many marked neuronal cells are present in each image. Aiming to automate the task for supporting research in the area, we propose a neural network architecture tuned specifically for this use case, cell ResUnet (c-ResUnet), and discuss the impact of alternative training strategies in overcoming particular challenges of our data. The approach provides good results in terms of both detection and counting, showing performance comparable to the interpretation of human operators. As a meaningful addition, we release the pre-trained model and the Fluorescent Neuronal Cells dataset collecting pixel-level annotations of where neuronal cells are located. In this way, we hope to help future research in the area and foster innovative methodologies for tackling similar problems. The second part deals with the problem of distributed data management in the context of LHC experiments, with a focus on supporting ATLAS operations concerning data transfer failures. In particular, we analyze error messages produced by failed transfers and propose a Machine Learning pipeline that leverages the word2vec language model and K-means clustering. This provides groups of similar errors that are presented to human operators as suggestions of potential issues to investigate. The approach is demonstrated on one full day of data, showing promising ability in understanding the message content and providing meaningful groupings, in line with previously reported incidents by human operators.

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.