Browsing by Author "Tampuu, Ardi, juhendaja"

Now showing 1 - 10 of 10

Collecting and Using a Labeled Dataset of NATO Mission Task Symbols to Improve and Benchmark Detection Models
(Tartu Ülikool, 2023) Açıkalın, Aral; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Neural networks are commonly used for object detection tasks but require immense amounts of data to train. For the task of North Atlantic Treaty Organization (NATO) mission task symbol detection using object detection neural networks, it is not possible to meet the data requirements. Additionally, labeling mission task symbols is very time-consuming and costly. This thesis aims to collect and label a dataset of NATO mission task symbols, propose a part of it as a benchmark for our solutions and future solutions, and finally propose different methods to use a part of the scarce collected data to improve the performance of our object detection models. YOLOv5 neural network is selected and used to experiment with different ways of using the scarce collected data. As a result, 113 images were collected and labeled. Five performance metrics are proposed for the benchmark. Finally, it was discovered that when dataset size is limited, extracting information from the dataset and using it to generate artificial data improves performance compared to directly introducing the scarce dataset to symbol detection models.
Comparing Output Modalities in End-to-End Driving
(Tartu Ülikool, 2022) Aidla, Romet; Matiisen, Tambet, juhendaja; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Self-driving car technology has made significant steps in the last ten years with the advancements in neural networks. The first autonomous vehicles are driving in San Francisco and Beijing. One of the promising approaches is end-to-end driving, where a neural network transforms an input image from a camera to output commands to control the vehicle. The most common output modalities are steering angle and trajectory. Both have been extensively benchmarked but not compared in similar settings. Metrics are usually calculated off-policy using a separated test dataset or on-policy using a simulator, but these have proven to correlate weakly with real-life performance. In this thesis, the comparison is made using an autonomous vehicle driving on WRC Rally Estonia tracks. The results show that the trajectory prediction approach is better at road positioning and recovering from non-ideal trajectories, which results in fewer situations where the safety driver has to take over.
Creation of Materials to Teach Data Science via Self-Driving
(Tartu Ülikool, 2024) Kreegipuu, Artur; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Andmeteaduse projektides esineb tihti kitsaskohti, millega õpilased võivad olla teoreetiliselt tuttavad, kuid puuduvad praktiliselt kogemused. Sageli esinevaid probleeme saab demonst-reerida isejuhtivate mudelautode abil. Praktiliste ülesannete kaudu saavad õpilased kogeda, kuidas isejuhtivate närvivõrkude arendamise etapis tehtavad andmeteaduslikud vead mõju-tavad isejuhtiva mudeli sooritusvõimet. Lõputöö raames loodi isejuhtivate mudelautode abil praktilised õppematerjalid eesmärgiga panna õpilased mõistma, ära tundma ja ennetama andmeteaduses laialdaselt levinud probleeme. Selleks loodi probleeme demonstreerivad praktilised ülesanded, mille käigus tuleb õpilastel koguda andmeid, treenida isejuhtivaid närvivõrke ja katsetada loodud mudeleid rajal. Kõik ülesanded lahendati testimise eesmärgil erinevates valgusoludes ja iga ülesande eeldatud tulemist filmiti video. Loodud õppemater-jalidele koguti tagasisidet kahelt masinõppe eksperdilt.
Dark diversity estimation based on a single matrix of binary observations
(Tartu Ülikool, 2020) Koger, Siim Karel; Tampuu, Ardi, juhendaja; Zafra, Raul Vicente, juhendaja; Pérez Carmona, Carlos, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Ecological theory and nature conservation have traditionally relied solely on observed local diversity. However, the biodiversity of a site also includes the absent species that are present in the surrounding region and can potentially inhabit the site’s particular ecological conditions. These unobserved species constitute the “dark diversity” of the site. Dark diversity is by definition unobservable and can only be estimated - in binary fashion or as a degree of certainty about species membership. This thesis compares the effectiveness of several implementations of non-negative matrix factorization (NMF) and basic autoencoders (a type of artificial neural network) to generate probabilistic values about a species’ membership in a specific site based on a single matrix of binary observations. We find that it is possible to generate a suitability matrix that is highly correlated with the underlying suitability values with both methods and that using autoencoders specifically for dark diversity predictions has a lot of potential for even more future improvements.
Driving Speed as a Hidden Factor Behind Distribution Shift
(Tartu Ülikool, 2022) Roosild, Kristjan; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Effect of Delays/Lag and Fighting it in Self-driving Neural Networks
(Tartu Ülikool, 2022) Uduste, Ilmar; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Autonomous driving is a field of interest for academia and industry alike, with hopes of fully replacing humans in the driver seat with artificial intelligence. Recent advances in the domain have been made in the use of end-to-end self-driving models as opposed to modular approaches. However, problem with building these end-to-end pipelines is that delays (lag) are commonly not taken into account. This work investigates the effect of lag in self-driving nets using Donkey Car, an autonomous car platform, and finds that the car drives worse when there is more lag in the pipeline. A novel method called frameshift is proposed to fight against the lag present and models using frameshift are proven to have significant real-life (closed-loop) performance gains over the baseline self-driving net, even in no-lag conditions and when comparing models analytically (open-loop). Frameshift is shown to be an effective tool in fighting against lag, although performance in very lag-heavy environments is inconsistent, as the car makes too few decisions per second to have consistent behaviour. The findings presented show the need to test self-driving cars in real-life (closedloop) as opposed to just analytically (open-loop) and also open up the field of end-to-end self-driving nets to the concept of using frameshift as a potential tool to fight against lag, although this novel idea requires further testing in different conditions and real-world data, to come to a final conclusion.
NATO standard mission task symbols pose detection based on symbol requirements
(Tartu Ülikool, 2023) Kõverik, Karl-Kristjan; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
The objective of this master’s thesis is to develop a series of neural network models that can accurately detect the pose of NATO mission task symbols. This research is significant as NATO mission task symbols play a critical role in planning military operations and automatically identifying their identity, location and pose can save valuable time and resources when digitizing military plans. The first step of research involved determining the essential components required to recreate mission task symbols into the KOLT digital planning system. This required a thorough analysis of the mission task symbols and their underlying characteristics. Through this analysis, we were able to identify the crucial features that the models must detect to sufficiently characterize each mission task symbol. The second task of research involves developing models that can detect the rotations of the mission task symbols accurately. Additionally, for trajectory symbols, it is essential to develop models that can detect and describe the symbols’ trajectory accurately. However, detecting poses of hand-drawn symbols is challenging, as there is often significant variability among symbols with the same meaning. Therefore, the model must be designed to account for this variability. The success of research will have significant implications for military operations. Accurate and efficient recognition of the poses of mission task symbols complements YOLO-based localization that is concurrently implemented in the research group, enabling AI-based digitalization of military plans. Fast digitalization enhances situational awareness and decision-making capabilities.
Prediction Models of Ischemic Stroke Using Deep Neural Networks
(Tartu Ülikool, 2021) Kurvits, Siim; Haller, Toomas, juhendaja; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
The ischemic stroke is one of the leading causes of death worldwide. Although, there are many known risk factors for the disease the growing amount of electronic medical data available gives opportunities for creating novel models for personal risk prediction. Usage of deep neural network (DNN) for developing such models can offers many benefits such as potential to encode multiple types of data, less feature selection and engineering required, and sometimes even an increased prediction accuracy. This Thesis focuses on developing a model for ischemic stroke prediction using electronic health record (EHR) data. I show that TabNet, a state-of-the art DNN architecture for tabular data analysis outperforms a simpler method, the FastAI tabular learner. Still, neither of the DNN methods achieved better results than the Random Forest. The ensemble models using Random Forest and DNN models were tested but only a small increase in the performance was achieved compared to the singular model. These results indicate that an ensemble-based methods such as Random Forest is sufficient for the data used. Nevertheless, with increased number of features and addition of more complex data types methods such as TabNet could still become valuable. All models developed resulted with high prediction power for ischemic stroke. This indicates that personal risk predictions for ischemic stroke can be given and the clinical utility of the models should be evaluated further.
Radial Softmax: A Novel Activation Function for Neural Networks to Reduce Overconfidence in Out-Of-Distribution Data
(Tartu Ülikool, 2020) Vagel, Rain; Tampuu, Ardi, juhendaja; Kull, Meelis, juhendaja; Vicente, Raul, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Neural networks are used widely and give state-of-the-art results in fields such as machine translation, image classification and speech recognition. These networks operate under the assumption that they predict on data that originates from the same distribution, as the training data. If this is not the case then the model will output incorrect results, often with very high confidence. In this work we explain how the commonly used softmax is unable to mitigate these problems and propose a new function called radial softmax which might help to mitigate out-of-distribution (OOD) overconfidence issues. We show that radial softmax is capable of mitigating OOD overconfidence issues in almost all cases. Based on our literature review this is the first time an improvement to softmax has been proposed for this issue. We also showed that changes to the training cycle or intermediate activation functions are not needed. With this function it is possible to make the models more resistant to OOD data without modifications to the larger architecture or training cycles. By having models that we know are resistant to OOD data, we can be more confident in the model output and use them for applications where mistakes are unacceptable such as healthcare, the defence industry or autonomous driving.
Testing Nvidia Drive for small cars in toy town
(Tartu Ülikool, 2022) Laanisto, Illimar; Tampuu, Ardi, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
The aim of this bachelor’s thesis is to analyze and evaluate the capabilities of neural network based image processing in Nvidia Drive in conditions far from its original operating domain. The data for the models is gathered from a toy town made of wood and populated with toy cars and toy pedestrians. This town is used for self-driving student competitions, teaching and public demonstrations. If applicable to this data, tools from NVIDIA Drive could be used for achieving self-driving with toy cars in this toy town, similar to how NVIDIA drive can be used on real cars in the real world. Three different Nvidia Drive models are tested in this thesis, which are DriveNet, PathNet, and OpenRoadNet. The performance of these models is evaluated and a conclusion is made.