Search

Computer vision systems that are deployed in safety-critical applications need to quantify their output uncertainty. We study regression from images to parameter values and here it is common to detect uncertainty by predicting probability distributions. In this context, we investigate the regression-by-classification paradigm which can represent multimodal distributions, without a prior assumption on the number of modes. Through experiments on a specifically designed synthetic dataset, we demonstrate that traditional loss functions lead to poor probability distribution estimates and severe overconfidence, in the absence of full ground truth distributions. In order to alleviate these issues, we propose hinge-Wasserstein – a simple improvement of the Wasserstein loss that reduces the penalty for weak secondary modes during training. This enables prediction of complex distributions with multiple modes, and allows training on datasets where full ground truth distributions are not available. In extensive experiments, we show that the proposed loss leads to substantially better uncertainty estimation on two challenging computer vision tasks: horizon line detection and stereo disparity estimation.

Machine learning is used today in a wide variety of applications, especially within computer vision, robotics, and autonomous systems. Example use cases include detecting people or other objects using cameras in autonomous vehicles, or navigating robots through collision-free paths to solve different tasks. The flexibility of machine learning is attractive as it can be applied to a wide variety of challenging tasks, without detailed prior knowledge of the problem domain. However, training machine learning models requires vast amounts of data, which leads to a significant manual effort, both for collecting the data and for annotating it. In this thesis, we study and develop methods for training machine learning models under in-sufficient data within computer vision, robotics, and autonomous systems, for the purpose of reducing the manual effort. In summary, we study (1) weakly-supervised learning for reducing the annotation cost, (2) methods for reducing model bias under highly imbalanced training data,(3) methods for obtaining trustworthy uncertainty estimates, and (4) the use of simulated and semi-virtual environments for reducing the amount of real-world data in reinforcement learning. In the first part of this thesis, we investigate how weakly-supervised learning can be used within image segmentation. In contrast to fully supervised learning, weakly-supervised learning uses a weaker form of annotation, which reduces the annotation effort. Typically, in image segmentation, each object needs to be precisely annotated in every image on the pixel level. Creating this type of annotation is both time consuming and costly. In weakly-supervised segmentation, however, the only information required is which objects are depicted in the images. This significantly reduces the annotation time. In Papers A and B, we propose two loss functions for improving the predicted object segmentations, especially their contours, in weakly-supervised segmentation. In the next part of th, Maskininlärning används idag i bred utsträckning inom många områden, och i synnerhet in-om datorseende, robotik, och autonoma system. Det kan till exempel användas för att detektera människor och andra föremål med kameror i autonoma bilar, eller för att styra robotar längs kollisionsfria banor för att lösa diverse uppgifter. Flexibiliteten i maskininlärning är attraktiv då den kan tillämpas för att lösa svåra problem utan detaljkännedom inom problemdomänen i fråga. Dock krävs en stor mängd data för att träna maskininlärningsmodeller, vilket medför en stor manuell arbetsbörda, dels för att samla in data, och dels för att annotera insamlade data. I denna avhandling undersöker och utvecklar vi metoder för att träna maskininlärningsmodeller med begränsad tillgång till data inom datorseende, robotik och autonoma system, i syfte att minska den manuella arbetsbördan. Sammanfattningsvis undersöker vi (1) svagt väglett läran-de för att minska annoteringstiden, (2) metoder som är opartiska under högt obalanserade data,(3) metoder för att erhålla pålitliga osäkerhetsskattningar, och (4) simulerings- och semivirtuella miljöer för att minska mängden riktiga data för förstärkningsinlärning. I den första delen av avhandlingen undersöker vi hur svagt väglett lärande (eng. weakly-supervised learning) kan användas inom bildsegmentering. Till skillnad från fullt väglett lärande används en svagare annoteringsform, vilket medför en minskning i den manuella annoterings-bördan. För bildsegmentering krävs i vanliga fall en noggrann annotering av varje enskilt objekt i varje bild på pixelnivå. Att skapa denna typ av annotering är både tidskrävande och kostsam. Med svagt väglett lärande krävs endast kännedom om vilka typer av objekt som finns i varje bild, vilket avsevärt minskar annoteringstiden. I Artikel A och B utformar vi två målfunktioner som är anpassade för att bättre segmentera objekt av intresse, i synnerhet deras konturer. I nästa del hanterar vi en oönskad effekt som kan uppstå u

Coverage path planning (CPP) is the problem of finding a path that covers the entire free space of a confined area, with applications ranging from robotic lawn mowing to search-and-rescue. When the environment is unknown, the path needs to be planned online while mapping the environment, which cannot be addressed by offline planning methods that do not allow for a flexible path space. We investigate how suitable reinforcement learning is for this challenging problem, and analyze the involved components required to efficiently learn coverage paths, such as action space, input feature representation, neural network architecture, and reward function. We propose a computationally feasible egocentric map representation based on frontiers, and a novel reward term based on total variation to promote complete coverage. Through extensive experiments, we show that our approach surpasses the performance of both previous RL-based approaches and highly specialized methods across multiple CPP variations., Funding agencies: y the Wallenberg AI, Autonomous Systems and Software Program (WASP), fundedby the Knut and Alice Wallenberg (KAW) Foundation; the Vinnova project, human centered autonomous regional airport, Dnr 2022-02678. The computational resources were provided by the National Academic Infrastructure for Supercomputing in Sweden (NAISS), partially funded by the Swedish Research Council through grant agreement no. 2022-06725, and by the Berzelius resource, provided by the KAW Foundation at the National Supercomputer Centre (NSC).

Image-level weakly-supervised semantic segmentation (WSSS) reduces the usually vast data annotation cost by surrogate segmentation masks during training. The typical approach involves training an image classification network using global average pooling (GAP) on convolutional feature maps. This enables the estimation of object locations based on class activation maps (CAMs), which identify the importance of image regions. The CAMs are then used to generate pseudo-labels, in the form of segmentation masks, to supervise a segmentation model in the absence of pixel-level ground truth. Our work is based on two techniques for improving CAMs; importance sampling, which is a substitute for GAP, and the feature similarity loss, which utilizes a heuristic that object contours almost always align with color edges in images. However, both are based on the multinomial posterior with softmax, and implicitly assume that classes are mutually exclusive, which turns out suboptimal in our experiments. Thus, we reformulate both techniques based on binomial posteriors of multiple independent binary problems. This has two benefits; their performance is improved and they become more general, resulting in an add-on method that can boost virtually any WSSS method. This is demonstrated on a wide variety of baselines on the PASCAL VOC dataset, improving the region similarity and contour quality of all implemented state-of-the-art methods. Experiments on the MS COCO dataset further show that our proposed add-on is well-suited for large-scale settings. Our code implementation is available at https://github.com/arvijj/hfpl.

Many real-world recognition problems are characterized by long-tailed label distributions. These distributions make representation learning highly challenging due to limited generalization over the tail classes. If the test distribution differs from the training distribution, e.g. uniform versus long-tailed, the problem of the distribution shift needs to be addressed. A recent line of work proposes learning multiple diverse experts to tackle this issue. Ensemble diversity is encouraged by various techniques, e.g. by specializing different experts in the head and the tail classes. In this work, we take an analytical approach and extend the notion of logit adjustment to ensembles to form a Balanced Product of Experts (BalPoE). BalPoE combines a family of experts with different test-time target distributions, generalizing several previous approaches. We show how to properly define these distributions and combine the experts in order to achieve unbiased predictions, by proving that the ensemble is Fisher-consistent for minimizing the balanced error. Our theoretical analysis shows that our balanced ensemble requires calibrated experts, which we achieve in practice using mixup. We conduct extensive experiments and our method obtains new state-of-the-art results on three long-tailed datasets: CIFAR-100-LT, ImageNet-LT, and iNaturalist-2018. Our code is available at https://github.com/emasa/BalPoE-CalibratedLT., Funding Agencies|Wallenberg Artificial Intelligence, Autonomous Systems and Software Program (WASP) - Knut and Alice Wallenberg Foundation; Swedish Research Council [2022-06725]; Knut and Alice Wallenberg Foundation at the National Supercomputer Centre