Modern deep neural networks are proficient at solving various visual recognition and understanding tasks, as long as a sufficiently large labeled dataset is available during the training time. However, the progress of these visual tasks is limited by the number of manual annotations. On the other hand, it is usually time-consuming and error-prone to annotate visual data, rendering the challenge of scaling up human labeling for many visual tasks. Fortunately, it is easy to collect large-scale, diverse unlabeled visual data from the Internet. And we can acquire a large amount of synthetic visual data with annotations from game engines effortlessly. In this dissertation, we explore how to utilize the unlabeled data and synthetic labeled data for various visual tasks, aiming to replace or reduce the direct supervision from the manual annotations. The key idea is to encourage deep neural networks to produce consistent predictions across different transformations (\eg geometry, temporal, photometric, etc.). We organize the dissertation as follows. In Part I, we propose to use the consistency over different geometric formulations and a cycle consistency over time to tackle the low-level scene geometry perception tasks in a self-supervised learning setting. In Part II, we tackle the high-level semantic understanding tasks in a semi-supervised learning setting, with the constraint that different augmented views of the same visual input maintain consistent semantic information. In Part III, we tackle the cross-domain image segmentation problem. By encouraging an adaptive segmentation model to output consistent results for a diverse set of strongly-augmented synthetic data, the model learns to perform test-time adaptation on unseen target domains with one single forward pass, without model training or optimization at the inference time. / Doctor of Philosophy / Recently, deep learning has emerged as one of the most powerful tools to solve various visual understanding tasks. However, the development of deep learning methods is significantly limited by the amount of manually labeled data. On the other hand, it is usually time-consuming and error-prone to annotate visual data, making the human labeling process not easily scalable. Fortunately, it is easy to collect large-scale, diverse raw visual data from the Internet (\eg search engines, YouTube, Instagram, etc.). And we can acquire a large amount of synthetic visual data with annotations from game engines effortlessly. In this dissertation, we explore how we can utilize the raw visual data and synthetic data for various visual tasks, aiming to replace or reduce the direct supervision from the manual annotations. The key idea behind this is to encourage deep neural networks to produce consistent predictions of the same visual input across different transformations (\eg geometry, temporal, photometric, etc.). We organize the dissertation as follows. In Part I, we propose using the consistency over different geometric formulations and a forward-backward cycle consistency over time to tackle the low-level scene geometry perception tasks, using unlabeled visual data only. In Part II, we tackle the high-level semantic understanding tasks using both a small amount of labeled data and a large amount of unlabeled data jointly, with the constraint that different augmented views of the same visual input maintain consistent semantic information. In Part III, we tackle the cross-domain image segmentation problem. By encouraging an adaptive segmentation model to output consistent results for a diverse set of strongly-augmented synthetic data, the model learns to perform test-time adaptation on unseen target domains.

Learning from unlabeled data is a topic of critical significance in machine learning, as the large datasets required to train ever-growing models are costly and impractical to annotate. Semi-Supervised Learning (SSL) methods aim to learn from a few labels and a large unlabeled dataset. In another approach, Domain Adaptation (DA) leverages data from a similar source domain to train a model for a target domain. This thesis focuses on Semi-Supervised Domain Adaptation (SSDA) for the dense task of semantic segmentation, where labels are particularly costly to obtain. SSDA has not received much attention yet, even though it has a great potential and represents a realistic scenario. The few existing SSDA methods for semantic segmentation reuse ideas from Unsupervised DA, despite the di↵erences between the two settings. This thesis proposes a new semantic segmentation framework designed particularly for the SSDA setting. The approach followed was to forego domain alignment and focus instead on enhancing clusterability of target domain features, an idea from SSL. The method is based on consistency regularization, combined with pixel contrastive learning and self-training. The proposed framework is found to be e↵ective not only in SSDA, but also in SSL. Ultimately, a unified solution for SSL and SSDA semantic segmentation is presented. Experiments were conducted on the target dataset of Cityscapes and source dataset of GTA5. The method proposed is competitive in both SSL and SSDA, and sets a new state-of-the-art for SSDA achieving a 65.6% mIoU (+4.4) on Cityscapes with 100 labeled samples. This thesis has an immediate impact on practical applications by proposing a new best-performing framework for the under-explored setting of SSDA. Furthermore, it also contributes towards the more ambitious goal of designing a unified solution for learning from unlabeled data. / Inlärning med hjälp av omärkt data är ett område av stor vikt inom maskininlärning. Detta på grund av att de stora datamängder som blivit nödvändiga för att träna konstant växande modeller både är kostsamma och opraktiska att implementera. Målet med Semi-Supervised Learning (SSL) är att kombinera ett fåtal etiketter med en stor mängd omärkt data för inlärning. Som ett annat tillvägagångssätt använder Domain Adaptation (DA) data från en liknande domän för att träna en annan måldomän. I Denna avhandling används Semi-Supervised Domain Adaptation (SSDA) för att utföra sådan semantisk segmentering, i vilken etiketter är särskilt kostsamma att erhålla. SSDA är ännu inte genererat mycket uppmärksamhet, även om det har en stor potential och representerar ett realistiskt scenario. De få metoder av SSDA som existerar för semantisk segmentering återanvänder idéer från Unsupervised DA, trots de olikheter som finns mellan de två modellerna. Denna avhandling föreslår ett nytt ramverk för semantisk segmentering, designat speciellt för SSDA modellen. Detta genom att försaka domänanpassning och i stället fokusera på att förbättra klusterbarheten av måldomänens egenskaper, en idé tagen från SSL. Metoden är baserad på konsistensregularisering, i kombination med pixelkontrastinlärning och självinlärning. Det föreslagna ramverket visar sig vara effektivt, inte bara för SSDA, men även för SSL. Till slut presenteras en enad lösning för semantisk segmentering med SLL och SSDA. Experiment utfördes på måldata från Cityscapes samt källdata från GTA5. Den föreslagna metoden är konkurrenskraftig både för SSL och SSDA, och blir världsledande för SSDA genom att uppnå 65,6% mIoU (+4,4) för Cityscapes med 100 märkta testdata. Denna avhandling har en omedelbar effekt gällande praktiska applikationer genom att föreslå ett nytt ”bäst resulterande” ramverk för dåligt utforskade inställningar av SSDA. Till yttermera visso bidrar avhandlingen även till det mer ambitiösa målet att designa en enad lösning för maskininlärning från omärkta data.