Abstract: Deep convolutional neural networks (CNNs) have achieved superior performance in many visual recognition application, such as image classification, detection and segmentation. In this thesis we address two limitations of CNNs. Training deep CNNs requires huge amounts of labeled data, which is expensive and labor intensive to collect. Another limitation is that training CNNs in a continual learning setting is still an open research question. Catastrophic forgetting is very likely when adapting trained models to new environments or new tasks. Therefore, in this thesis, we aim to improve CNNs for applications with limited data and to adapt CNNs continually to new tasks.
Self-supervised learning leverages unlabelled data by introducing an auxiliary task for which data is abundantly available. In the first part of the thesis, we show how rankings can be used as a proxy self-supervised task for regression problems. Then we propose an efficient backpropagation technique for Siamese networks which prevents the redundant computation introduced by the multi-branch network architecture. In addition, we show that measuring network uncertainty on the self-supervised proxy task is a good measure of informativeness of unlabeled data. This can be used to drive an algorithm for active learning. We then apply our framework on two regression problems: Image Quality Assessment (IQA) and Crowd Counting. For both, we show how to automatically generate ranked image sets from unlabeled data. Our results show that networks trained to regress to the ground truth targets for labeled data and to simultaneously learn to rank unlabeled data obtain significantly better, state-of-the-art results. We further show that active learning using rankings can reduce labeling effort by up to 50\% for both IQA and crowd counting.
In the second part of the thesis, we propose two approaches to avoiding catastrophic forgetting in sequential task learning scenarios. The first approach is derived from Elastic Weight Consolidation, which uses a diagonal Fisher Information Matrix (FIM) to measure the importance of the parameters of the network. However the diagonal assumption is unrealistic. Therefore, we approximately diagonalize the FIM using a set of factorized rotation parameters. This leads to significantly better performance on continual learning of sequential tasks. For the second approach, we show that forgetting manifests differently at different layers in the network and propose a hybrid approach where distillation is used in the feature extractor and replay in the classifier via feature generation. Our method addresses the limitations of generative image replay and probability distillation (i.e. learning without forgetting) and can naturally aggregate new tasks in a single, well-calibrated classifier. Experiments confirm that our proposed approach outperforms the baselines and some start-of-the-art methods.

Abstract: An improved method for visual gamma correction is developed for LCD displays to increase the accuracy of digital colour reproduction. Rather than utilising a photometric measurement device, we use observ- ers’ visual luminance judgements for gamma correction. Eight half tone patterns were designed to gen- erate relative luminances from 1/9 to 8/9 for each colour channel. A psychophysical experiment was conducted on an LCD display to find the digital signals corresponding to each relative luminance by visually matching the half-tone background to a uniform colour patch. Both inter- and intra-observer vari- ability for the eight luminance matches in each channel were assessed and the luminance matches proved to be consistent across observers (DE00 < 3.5) and repeatable (DE00 < 2.2). Based on the individual observer judgements, the display opto-electronic transfer function (OETF) was estimated by using either a 3rd order polynomial regression or linear interpolation for each colour channel. The performance of the proposed method is evaluated by predicting the CIE tristimulus values of a set of coloured patches (using the observer-based OETFs) and comparing them to the expected CIE tristimulus values (using the OETF obtained from spectro-radiometric luminance measurements). The resulting colour differences range from 2 to 4.6 DE00. We conclude that this observer-based method of visual gamma correction is useful to estimate the OETF for LCD displays. Its major advantage is that no particular functional relationship between digital inputs and luminance outputs has to be assumed.

Abstract: arXiv:1706.01487
In this paper we propose an approach to lexicon-free recognition of text in scene images. Our approach relies on a LSTM-based soft visual attention model learned from convolutional features. A set of feature vectors are derived from an intermediate convolutional layer corresponding to different areas of the image. This permits encoding of spatial information into the image representation. In this way, the framework is able to learn how to selectively focus on different parts of the image. At every time step the recognizer emits one character using a weighted combination of the convolutional feature vectors according to the learned attention model. Training can be done end-to-end using only word level annotations. In addition, we show that modifying the beam search algorithm by integrating an explicit language model leads to significantly better recognition results. We validate the performance of our approach on standard SVT and ICDAR'03 scene text datasets, showing state-of-the-art performance in unconstrained text recognition.

Abstract: TB6.2
Road segmentation is an essential functionality for supporting advanced driver assistance systems (ADAS) such as road following and vehicle and pedestrian detection. Significant efforts have been made in order to solve this task using vision-based techniques. The major challenge is to deal with lighting variations and the presence of objects on the road surface. In this paper, we propose a new road detection method to infer the areas of the image depicting road surfaces without performing any image segmentation. The idea is to previously segment manually or semi-automatically the road region in a traffic-free reference video record on a first drive. And then to transfer these regions to the frames of a second video sequence acquired later in a second drive through the same road, in an on-line manner. This is possible because we are able to automatically align the two videos in time and space, that is, to synchronize them and warp each frame of the first video to its corresponding frame in the second one. The geometric transform can thus transfer the road region to the present frame on-line. In order to reduce the different lighting conditions which are present in outdoor scenarios, our approach incorporates a shadowless feature space which represents an image in an illuminant-invariant feature space. Furthermore, we propose a dynamic background subtraction algorithm which removes the regions containing vehicles in the observed frames which are within the transferred road region.

Abstract: Autonomous driving is a key factor for future mobility. Properly perceiving the environment of the vehicles is essential for a safe driving, which requires computing accurate geometric and semantic information in real-time. In this paper, we challenge state-of-the-art computer vision algorithms for building a perception system for autonomous driving. An inherent drawback in the computation of visual semantics is the trade-off between accuracy and computational cost. We propose to circumvent this problem by following an offline-online strategy. During the offline stage dense 3D semantic maps are created. In the online stage the current driving area is recognized in the maps via a re-localization process, which allows to retrieve the pre-computed accurate semantics and 3D geometry in realtime. Then, detecting the dynamic obstacles we obtain a rich understanding of the current scene. We evaluate quantitatively our proposal in the KITTI dataset and discuss the related open challenges for the computer vision community.

Abstract: The manuscript evaluates the performance of a monocular visual odometry approach when images from different spectra are considered, both independently and fused. The objective behind this evaluation is to analyze if classical approaches can be improved when the given images, which are from different spectra, are fused and represented in new domains. The images in these new domains should have some of the following properties: i) more robust to noisy data; ii) less sensitive to changes (e.g., lighting); iii) more rich in descriptive information, among other. In particular in the current work two different image fusion strategies are considered. Firstly, images from the visible and thermal spectrum are fused using a Discrete Wavelet Transform (DWT) approach. Secondly, a monochrome threshold strategy is considered. The obtained
representations are evaluated under a visual odometry framework, highlighting
their advantages and disadvantages, using different urban and semi-urban scenarios. Comparisons with both monocular-visible spectrum and monocular-infrared spectrum, are also provided showing the validity of the proposed approach.