To design more efficient sailing boat sails and to analyze the efficiency of a sail trim on the water, it is very helpful to have the ability to obtain a digital copy of real-life sail configurations. As a step towards obtaining such digital copies, the Sailing Innovation Centre in collaboration with GeoDelta has created point cloud measurements from sails by placing physical markers on them. This paper introduces a physics-based reconstruction method that uses a known sail rest shape combined with this point cloud data to reconstruct the sail configuration at the time that the measurement was taken. By using a physics-based sail model and the knowledge that the sail is subject to gravity and wind forces together with a known sail rest shape, we can create plausible reconstructions even for areas in the sail where fewer to no markers are placed. Using measurements that were taken before this reconstruction method was proposed in combination with an approximated sail rest shape and material properties, our reconstruction method was approximated to have an average reconstruction distance below 1cm for a 2.5m by 4.5m sail. Taking new measurements following our proposed guidelines in combination with having a more correct sail rest shape and better knowing the sail material properties is very likely to yield even better results.

Event-based cameras do not capture frames like an RGB camera, only data from pixels that detect a change in light intensity, making it a better alternative for processing videos. The sparse data acquired from event-based video only captures movement in an asynchronous way. In this paper an evaluation is made on the efficiency and accuracy of object detection, specifically localization, between sparse and dense representations of data. Convolutional Neural Networks are used to train and test on images and event-based data. The results show a positive trade-off in terms of accuracy and efficiency for using sparse event-based data instead of dense data like images. These results provide a basis for an argument to use event-based cameras instead of RGB cameras when dealing with object detection.

Event-based cameras represent a new alternative to traditional frame based sensors, with advantages in lower output bandwidth, lower latency and higher dynamic range, thanks to their independent, asynchronous pixels. These advantages prompted the development of computer vision methods on event data in the last decade, however event-based datasets are still in early stages in terms of size and complexity compared to normal datasets (e.g. ImageNet). This paper explores event data augmentation by superimposing two existing event datasets (N-MNIST and N-Caltech101) and by adding uniform noise. It shows that training an instance segmentation model on noisy datasets does not improve its performance, but the amount and type of noise added in the background decreases the performance of such model.

Instance segmentation on data from Dynamic Vision Sensors (DVS) is an important computer vision task that needs to be tackled in order to push the research forward on these types of inputs. This paper aims to show that deep learning based techniques can be used to solve the task of instance segmentation on DVS data. A high performing model was used to solve this task, using event-based data that was transformed into RGB-D images. The chosen model for this work was Mask R-CNN, with an alteration for depth images, because of its high performance on frame based data. The N-MNIST dataset provides the event-based input, and the transformation of such an input is presented in this study. Furthermore, the masks are generated with the help of the MNIST dataset and heuristics are used for placing them at the correct positions. The results are promising and comparable to other results from literature on the task of semantic segmentation.

The event-based camera represents a revolutionary concept, having an asynchronous output. The pixels of dynamic vision sensors react to the brightness change, resulting in streams of events at very small intervals of time. This paper provides a model to track objects in neuromorphic datasets, using clustering. In addition, a non-linear filter is applied to correct the estimation of the object position. Both single and multi-object tracking algorithms are provided and their performance is analyzed using different metrics, including the clustering evaluation scores and the tracking accuracy. The accuracy is over 0.6 for multi-target tracking and more than 0.7 for single object tracking. Besides the proposed model, a comparison between different possible approaches for event-based data tracking is provided.

AI systems have the ability to complete tasks with greater precision and speed than humans, which has led to an increase in their usage. These systems are often grouped with humans in order to take advantage of the unique abilities of both the AI and the human. However, to make this cooperation as efficient as possible, there needs to be a mutual trust between humans and AIs. While there has been much research concerning the topic of human trust, there is a lack of work done concerning the trust that an artificial agent has toward its human partner. Given that a human must appear trustworthy to an artificial agent in order for that agent to trust him, and that demanding and offering help are important parts of a collaboration, the following research question has been formulated : How does an artificial agent asking a human for advice or help affect that human's trustworthiness ? To answer this research question, an experiment was conducted through an urban search and rescue game using the MATRX Software. Through this game, participants had to collaborate with a robot partner in order to accomplish the task of finding and rescuing 8 victims and delivering them to the correct drop zone. The participants were divided into a control group, who worked alongside a basic rescue robot, and an experimental group, which had a help-seeker robot as a partner. The help-seeker robot differed from the basic robot in its ability to ask the participant for advice, such as asking which room it should search for victims to rescue. Following the experiment, no significant results indicating a positive or negative effect on human trustworthiness by the help-seeker agent's behaviour were found.

The number of collaborations between humans and artificial agents has risen steeply in recent years due to the rapid expansion of AI. Numerous studies in social sciences have already established that trust is a crucial factor in ensuring effective teamwork. While the dynamics of trust in human-human relationships or the effects of human comportment on the artificial agents' trustworthiness have been researched, the changes in human trustworthiness as a result of different AI behaviors were scarcely analysed. Therefore, the latter perspective needs to be thoroughly studied as well. This paper has investigated how an AI directing the human can affect human trustworthiness. So far, the directability has been studied as a notion on its own without being linked to trust or its effects on human-AI relationships. For this research, subjective and objective measurements have been used to assess the effects of an AI directing the human in joint activity. In order to quantify the human trustworthiness and assess the possible changes, an experiment has been carried out where participants played a Search & Rescue game developed in the MATRX software. The human's trustworthiness has been calculated using the ABI model and a questionnaire. The obtained results suggest that there is not a significant change in human trustworthiness as a result of the artificial agent directing the human.

Collaborative AI (CAI) is a fast growing field of study. Cooperation between teams composed of humans and artificial intelligence needs to be principled and founded on reciprocal trust. Modelling the trustworthiness of humans is a difficult task because of the ambiguous nature of its definition as well as the effect of team work dynamic. This research defines and measures human trustworthiness within the context of human-AI collaboration and tests how the artificial intelligence agent’s action of offering help plays a role in it. The experiment is conducted through the MATRX framework, in which a human agent and an AI agent collaborate in order to search and rescue victims within the environment. The ABI trust model is used to determine the sub-components that define trustworthiness, which is ability, benevolence and integrity. Trustworthiness is measured in 2 ways, through objective measures which represent the AI agent’s measure of the human trustworthiness and subjective measures that represent the human’s measure of their own trustworthiness. The results show that help offered by the AI agent, improves the ability and benevolence of the human agent in objective metrics but not integrity. Subjective results show no statistically significant change. The research concludes that the trustworthiness perceived by the AI agent is indeed improved, but does not provide evidence of the same from the perception of the human agent.

As technology advances, automated systems become more autonomous which leads to a higher interdependence between machine and human. Much research has been done about trust between humans and trust of humans regarding machines. An interesting question that remains is how the behavior of an agent influences human trustworthiness in a human-agent collaborative setting. The research presented by this paper contributes to the understanding of this area. It investigates a specific behavioral trait using the following hypothesis: friendly behavior of an agent improves human trustworthiness. Here, trustworthiness is broken up in the constructs: ability, benevolence and integrity.

An experiment has been conducted using a collaborative Search and Rescue game. The following behaviors of the participants have been measured:
- Ability: speed and effectiveness;
- Benevolence: communication, willingness to help, agreeableness to advice, responsiveness;
- Integrity: truthfulness.
Furthermore, a likert scale has been used to measure the participants' own perception of their trustworthiness. The experiment is conducted with 20 participants in the control group, where the agent spoke in a neutral manner, and 20 in the experimental group, where the agent instilled empathy, stimulated collaboration, encouraged the participants and was affectionate.

The research has shown a significant improvement in the experimental group only for communication and willingness to help. This gives some indication that a friendly agent only slightly improves the trustworthiness of a human. However, the research has some limitations that might also explain the lack of significant results. Firstly, it is unclear to what extent the measures truly measured the constructs of trustworthiness. Secondly, to create a friendly agent, theories from organizational and social psychology are used, which are mostly focussed on human-human relationships, instead of human-agent relationships. Finally, Some confounding variables may have had an impact, like lag in the game and the participant not properly reading the agent’s messages.

Human-AI teams require trust to operate efficiently and solve certain tasks like search & rescue. Trustworthiness is measured using the ABI model; Ability, Benevolence and Integrity. This research paper tries to observe the effect a conflicting robot has on the human trustworthiness. The hypothesis we try to test is: “human trustworthiness will decrease when paired with a conflicting AI”. We conduct an experiment with one control group playing with a normal agent and an experiment group paired with the conflicting agent. Using the ABI concepts, we model the human trustworthiness across both groups using in-game observations (objective) and a questionnaire (subjective). When comparing the results from both group we see that the conflicting agent does not decrease the objective trustworthiness, however looking at the questionnaires we observe that the subjective human benevolence and integrity are negatively affected when paired with the conflicting agent.

Convolutional neural networks (CNNs) are often pruned to achieve faster training and inference speed while also requiring less memory. Nevertheless, during computation, most modern GPUs cannot take advantage of the sparsity automatically, especially on networks with unstructured sparsity. Therefore, many libraries that exploit sparsity, have been proposed for accelerating CNN inference on GPUs. However, there is little research on systematically comparing them. In this paper, some state-of-the-art libraries for accelerating sparse CNN inference on GPUs are reviewed and benchmarked. Most of the libraries speedup the convolution and/or pooling operations by skipping zero calculations, therefore, they are able to perform sparse matrix calculations faster. However, many of them have hardware and software restrictions and are hard to integrate into a new model to perform end-to-end inference.

Color information has been shown to provide useful information during image classification. Yet current popular deep convolutional neural networks use 2-dimensional convolutional layers. The first 2-dimensional convolutional layer in the network combines the color channels of the input images, which produces feature maps per channel with only spatial dimensions, height and width, getting rid of the color dimension. In this work we introduce Full Color Deep networks which use 3-dimensional convolutions to retain the color dimension beyond the first layer. The 3D kernels convolve over the color and spatial dimensions. The network can extract features from all three dimensions in all layers which are subsequently used by the classifier. We show that the Full Color Deep networks perform at least as well as the current CNNs but outperform them in learning color information and using that information in other downstream tasks.

In recent years, the expansion of the Internet has brought an explosion of visual information, including social media, medical photographs, and digital history. This massive amount of visual content generation and sharing presents new challenges, especially when searching for similar information in databases —— Content-Based Image Retrieval (CBIR). Feature extraction is the foundation of image retrieval, making research into obtaining concrete features and representations of image content a vital concern.
In the feature extraction module, We first pre-process the target image and input it into a CNN to obtain feature maps for different channels. These feature maps can be aggregated into compact and global uniform descriptors by pooling. Then these global descriptors are further dimensionalised and normalized by whitening methods to obtain image feature vectors that are easy to compute and compare. In this process, the accuracy of the retrieval depends on how accurately the final feature vectors represent the meaning expressed by the target image. Therefore, various CNN network structures, pooling and whitening methods are proposed to get more concrete feature vectors.
In this thesis, our study (1) fine tunes the pre-trained CNNs, (2) optimizes the application of second-order attention information in feature map, (3) applies and compares popular feature enhancement methods in both aggregating and whitening, (4) explores how to combine all strengths, and (5) propose a new model \textit{ResNet-SOI}, which achieves 53.4(M) and 59.2(M) mAP on the challenging benchmark \textit{ROxford5k+1M} and\textit{ RParis6k+1M}, and outperforms the state-of-art methods. Our prototype GUI is available on GitHub (https://github.com/yanan-huu/Image-Search-Engine-for-Historical-Research).

Deep learning has been widely implemented in industrial inspection, such as damage detection from images. However, training deep networks requires massive data, which is hard to collect and laborious to annotate, especially in the aviation scenario of aircraft engines. To alleviate the demand for annotated data, we create BladeSynth - a large synthetic image dataset for detecting damage from aircraft engines, and empirically evaluate the transferability of state-of-the-art Scaled-YOLOV4 from synthetic to real world by pre-training on synthetic data and fine-tuning on real data. Our experiments show that pre-training on synthetic data improves the performance in damage detection in aircraft engine images.

In this paper we analyze the performance of a novel clustering objective that optimizes a neural network to predict segmentation. We challenge the reported results by replicating the original experiments and conducting additional tests to gain an insight into the algorithm. We analyzed the efficiency of the clustering objective on a different architecture, dataset and hyper-parameters. To our surprise the algorithm demonstrated considerably lower results when running on the new setup. Further, in our work we detail the reasons behind the discrepancy and provide configurations under which the method performs best. We show that the objective is highly sensitive to the type of images it is predicting and the complexity of the architecture that is being used with.

Convolutional Neural Networks are particularly vulnerable to attacks that manipulate their data, which are usually called adversarial attacks. In this paper, a method of filtering images using the Fast Fourier Transform is explored, along with its potential to be used as a defense mechanism to such attacks. The main contribution that differs from other methods that use the Fourier Transform as a filtering element in neural networks is the use of labeled data to determine how to filter the pictures. This paper concludes that, while the filtering proposed is hardly better than a simple low-pass filter, it still manages to improve resistance to adversarial attacks with a minimal drop in the standard accuracy of the network.

A structured CNN filter basis allows incorporating priors about natural image statistics and thus require less training examples to learn, saving valuable annotation time. Here, we build on the Gaussian derivative CNN filter basis that learn both the orientation and scale of the filters. However, this Gaussian filter basis definition depends on a predetermined derivative order, which typically results in fixed frequency responses for the basis functions whereas the optimal frequency of the filters should depend on the data and the downstream learning task. We show that by learning the order of the basis we can accurately learn the frequency of the filters, and hence adapt to the optimal frequencies for the underlying task. We investigate the well-founded mathematical formulation of fractional derivatives to adapt the filter frequencies during training. Our formulation leads to parameter savings and data efficiency when compared to the standard CNNs and the Gaussian derivative CNN filter networks that we build on.