Abstract: Color description is one of the fundamental problems of image understanding. One of the popular ways to represent colors is by means of color names. Most existing work on color names focuses on only the eleven basic color terms of the English language. This could be limiting the discriminative power of these representations, and representations based on more color names are expected to perform better. However, there exists no clear strategy to choose additional color names. We collect a dataset of 28 additional color names. To ensure that the resulting color representation has high discriminative power we propose a method to order the additional color names according to their complementary nature with the basic color names. This allows us to compute color name representations with high discriminative power of arbitrary length. In the experiments we show that these new color name descriptors outperform the existing color name descriptor on the task of visual tracking, person re-identification and image classification.

Abstract: One of the fundamental problems of computer vision is to represent images with compact semantically relevant embeddings. These embeddings could then be used in a wide variety of applications, such as image retrieval, object detection, and video search. The main objective of this thesis is to study image embeddings from two aspects: color embeddings and deep embeddings.
In the first part of the thesis we start from hand-crafted color embeddings. We propose a method to order the additional color names according to their complementary nature with the basic eleven color names. This allows us to compute color name representations with high discriminative power of arbitrary length. Psychophysical experiments confirm that our proposed method outperforms baseline approaches. Secondly, we learn deep color embeddings from weakly labeled data by adding an attention strategy. The attention branch is able to correctly identify the relevant regions for each class. The advantage of our approach is that it can learn color names for specific domains for which no pixel-wise labels exists.
In the second part of the thesis, we focus on deep embeddings. Firstly, we address the problem of compressing large embedding networks into small networks, while maintaining similar performance. We propose to distillate the metrics from a teacher network to a student network. Two new losses are introduced to model the communication of a deep teacher network to a small student network: one based on an absolute teacher, where the student aims to produce the same embeddings as the teacher, and one based on a relative teacher, where the distances between pairs of data points is communicated from the teacher to the student. In addition, various aspects of distillation have been investigated for embeddings, including hint and attention layers, semi-supervised learning and cross quality distillation. Finally, another aspect of deep metric learning, namely lifelong learning, is studied. We observed some drift occurs during training of new tasks for metric learning. A method to estimate the semantic drift based on the drift which is experienced by data of the current task during its training is introduced. Having this estimation, previous tasks can be compensated for this drift, thereby improving their performance. Furthermore, we show that embedding networks suffer significantly less from catastrophic forgetting compared to classification networks when learning new tasks.

Abstract: In this work we contribute a novel pipeline to automatically generate training data, and to improve over state-of-the-art multi-object tracking and segmentation (MOTS) methods. Our proposed track mining algorithm turns raw street-level videos into high-fidelity MOTS training data, is scalable and overcomes the need of expensive and time-consuming manual annotation approaches. We leverage state-of-the-art instance segmentation results in combination with optical flow predictions, also trained on automatically harvested training data. Our second major contribution is MOTSNet – a deep learning, tracking-by-detection architecture for MOTS – deploying a novel mask-pooling layer for improved object association over time. Training MOTSNet with our automatically extracted data leads to significantly improved sMOTSA scores on the novel KITTI MOTS dataset (+1.9%/+7.5% on cars/pedestrians), and MOTSNet improves by +4.1% over previously best methods on the MOTSChallenge dataset. Our most impressive finding is that we can improve over previous best-performing works, even in complete absence of manually annotated MOTS training data.

Abstract: Nowadays, automatic identity document recognition, including passport and driving license recognition, is at the core of many applications within the administrative and service sectors, such as police, hospitality, car renting, etc. In former years, the document information was manually extracted whereas today this data is recognized automatically from images obtained by flat-bed scanners. Yet, since these scanners tend to be expensive and voluminous, companies in the sector have recently turned their attention to cheaper, small and yet computationally powerful scanners: the mobile devices. The document identity recognition from mobile images enclose several new difficulties w.r.t traditional scanned images, such as the loss of a controlled background, perspective, blurring, etc. In this paper we present a real application for identity document classification of images taken from mobile devices. This classification process is of extreme importance since a prior knowledge of the document type and origin strongly facilitates the subsequent information extraction. The proposed method is based on a traditional Bagof-Words in which we have taken into consideration several key aspects to enhance recognition rate. The method performance has been studied on three datasets containing more than 2000 images from 129 different document classes.

Abstract: In this paper we present a knowledge base of architectural documents aiming at improving existing methods of floor plan classification and understanding. It consists of an ontological definition of the domain and the inclusion of real instances coming from both, automatically interpreted and manually labeled documents. The knowledge base has proven to be an effective tool to structure our knowledge and to easily maintain and upgrade it. Moreover, it is an appropriate means to automatically check the consistency of relational data and a convenient complement of hard-coded knowledge interpretation systems.

Abstract: Architectural floorplans exhibit a large variability in notation. Therefore, segmenting and identifying the elements of any kind of plan becomes a challenging task for approaches based on grouping structural primitives obtained by vectorization. Recently, a patch-based segmentation method working at pixel level and relying on the construction of a visual vocabulary has been proposed showing its adaptability to different notations by automatically learning the visual appearance of the elements in each different notation. In this paper we describe an evolution of this new approach in two directions: firstly we evaluate different features to obtain the description of every patch. Secondly, we train an SVM classifier to obtain the category of every patch instead of constructing a visual vocabulary. These modifications of the method have been tested for wall detection on two datasets of architectural floorplans with different notations and compared with the results obtained with the original approach.

Abstract: This paper presents a syntactic model for architectural floor plan interpretation. A stochastic image grammar over an And-Or graph is inferred to represent the hierarchical, structural and semantic relations between elements of all possible floor plans. This grammar is augmented with three different probabilistic models, learnt from a training set, to account the frequency of that relations. Then, a Bottom-Up/Top-Down parser with a pruning strategy has been used for floor plan recognition. For a given input, the parser generates the most probable parse graph for that document. This graph not only contains the structural and semantic relations of its elements, but also its hierarchical composition, that allows to interpret the floor plan at different levels of abstraction.