Abstract: In this dissertation, we explore different ways to perform human action recognition in video clips. We focus on data efficiency, proposing new approaches that alleviate the need for laborious and time-consuming manual data annotation. In the first part of this dissertation, we start by analyzing previous state-of-the-art models, comparing their differences and similarities in order to pinpoint where their real strengths come from. Leveraging this information, we then proceed to boost the classification accuracy of shallow models to levels that rival deep neural networks. We introduce hybrid video classification architectures based on carefully designed unsupervised representations of handcrafted spatiotemporal features classified by supervised deep networks. We show in our experiments that our hybrid model combine the best of both worlds: it is data efficient (trained on 150 to 10,000 short clips) and yet improved significantly on the state of the art, including deep models trained on millions of manually labeled images and videos. In the second part of this research, we investigate the generation of synthetic training data for action recognition, as it has recently shown promising results for a variety of other computer vision tasks. We propose an interpretable parametric generative model of human action videos that relies on procedural generation and other computer graphics techniques of modern game engines. We generate a diverse, realistic, and physically plausible dataset of human action videos, called PHAV for “Procedural Human Action Videos”. It contains a total of 39,982 videos, with more than 1,000 examples for each action of 35 categories. Our approach is not limited to existing motion capture sequences, and we procedurally define 14 synthetic actions. We then introduce deep multi-task representation learning architectures to mix synthetic and real videos, even if the action categories differ. Our experiments on the UCF-101 and HMDB-51 benchmarks suggest that combining our large set of synthetic videos with small real-world datasets can boost recognition performance, outperforming fine-tuning state-of-the-art unsupervised generative models of videos.

Abstract: The image sensor, nowadays, is rolling the smartphone industry. While some phone brands explore equipping more image sensors, others, like Google, maintain their smartphones with just one sensor; but this sensor is equipped with Deep Learning to enhance the image quality. However, what all brands agree on is the need to research new image sensors; for instance, in 2015 Omnivision and PixelTeq presented new CMOS based image sensors defined as multispectral Single Sensor Camera (SSC), which are capable of capturing multispectral bands. This dissertation presents the benefits of using a multispectral SSCs that, as aforementioned, simultaneously acquires images in the visible and near-infrared (NIR) bands. The principal benefits while addressing problems related to image bands in the spectral range of 400 to 1100 nanometers, there are cost reductions in the hardware and software setup because only one SSC is needed instead of two, and the images alignment are not required any more. Concerning to the NIR spectrum, many works in literature have proven the benefits of working with NIR to enhance RGB images (e.g., image enhancement, remove shadows, dehazing, etc.). In spite of the advantage of using SSC (e.g., low latency), there are some drawback to be solved. One of this drawback corresponds to the nature of the silicon-based sensor, which in addition to capture the RGB image, when the infrared cut off filter is not installed it also acquires NIR information into the visible image. This phenomenon is called RGB and NIR crosstalking. This thesis firstly faces this problem in challenging images and then it shows the benefit of using multispectral images in the edge detection task.
The RGB color restoration from RGBN image is the topic tackled in RGB and NIR crosstalking. Even though in the literature a set of processes have been proposed to face this issue, in this thesis novel approaches, based on DL, are proposed to subtract the additional NIR included in the RGB channel. More precisely, an Artificial Neural Network (NN) and two Convolutional Neural Network (CNN) models are proposed. As the DL based models need a dataset with a large collection of image pairs, a large dataset is collected to address the color restoration. The collected images are from challenging scenes where the sunlight radiation is sufficient to give absorption/reflectance properties to the considered scenes. An extensive evaluation has been conducted on the CNN models, differences from most of the restored images are almost imperceptible to the human eye. The next proposal of the thesis is the validation of the usage of SSC images in the edge detection task. Three methods based on CNN have been proposed. While the first one is based on the most used model, holistically-nested edge detection (HED) termed as multispectral HED (MS-HED), the other two have been proposed observing the drawbacks of MS-HED. These two novel architectures have been designed from scratch (training from scratch); after the first architecture is validated in the visible domain a slight redesign is proposed to tackle the multispectral domain. Again, another dataset is collected to face this problem with SSCs. Even though edge detection is confronted in the multispectral domain, its qualitative and quantitative evaluation demonstrates the generalization in other datasets used for edge detection, improving state-of-the-art results.

Abstract: Robust text reading from street view images provides valuable information for various applications. Performance improvement of existing methods in such a challenging scenario heavily relies on the amount of fully annotated training data, which is costly and in-efficient to obtain. To scale up the amount of training data while keeping the labeling procedure cost-effective, this competition introduces a new challenge on Large-scale Street View Text with Partial Labeling (LSVT), providing 50, 000 and 400, 000 images in full and weak annotations, respectively. This competition aims to explore the abilities of state-of-the-art methods to detect and recognize text instances from large-scale street view images, closing the gap between research benchmarks and real applications. During the competition period, a total of 41 teams participated in the two proposed tasks with 132 valid submissions, ie, text detection and end-to-end text spotting. This paper includes dataset descriptions, task definitions, evaluation protocols and results summaries of the ICDAR 2019-LSVT challenge.

Abstract: While many individual tasks in the domain of human analysis have recently received an accuracy boost from deep learning approaches, multi-task learning has mostly been ignored due to a lack of data. New synthetic datasets are being released, filling this gap with synthetic generated data. In this work, we analyze four related human analysis tasks in still images in a multi-task scenario by leveraging such datasets. Specifically, we study the correlation of 2D/3D pose estimation, body part segmentation and full-body depth estimation. These tasks are learned via the well-known Stacked Hourglass module such that each of the task-specific streams shares information with the others. The main goal is to analyze how training together these four related tasks can benefit each individual task for a better generalization. Results on the newly released SURREAL dataset show that all four tasks benefit from the multi-task approach, but with different combinations of tasks: while combining all four tasks improves 2D pose estimation the most, 2D pose improves neither 3D pose nor full-body depth estimation. On the other hand 2D parts segmentation can benefit from 2D pose but not from 3D pose. In all cases, as expected, the maximum improvement is achieved on those human body parts that show more variability in terms of spatial distribution, appearance and shape, e.g. wrists and ankles.

Abstract: Most high-level computer vision tasks rely on low-level image operations as their initial processes. Operations such as edge detection, image enhancement, and super-resolution, provide the foundations for higher level image analysis. In this work we address the edge detection considering three main objectives: simplicity, efficiency, and generalization since current state-of-the-art (SOTA) edge detection models are increased in complexity for better accuracy. To achieve this, we present Tiny and Efficient Edge Detector (TEED), a light convolutional neural network with only 58K parameters, less than 0:2% of the state-of-the-art models. Training on the BIPED dataset takes less than 30 minutes, with each epoch requiring less than 5 minutes. Our proposed model is easy to train and it quickly converges within very first few epochs, while the predicted edge-maps are crisp and of high quality. Additionally, we propose a new dataset to test the generalization of edge detection, which comprises samples from popular images used in edge detection and image segmentation. The source code is available in https://github.com/xavysp/TEED.

Abstract: This article argues that 2-alternative forced choice tasks and Thurstone’s law of comparative judgments (Thurstone, 1927) are well suited to investigate code-switching competence by means of acceptability judgments. We compare this method with commonly used Likert scale judgments and find that the 2-alternative forced choice task provides granular details that remain invisible in a Likert scale experiment. In order to compare and contrast both methods, we examined the syntactic phenomenon usually referred to as the Adjacency Condition (AC) (apud Stowell, 1981), which imposes a condition of adjacency between verb and object. Our interest in the AC comes from the fact that it is a subtle feature of English grammar which is absent in Spanish, and this provides an excellent springboard to create minimal code-switched pairs that allow us to formulate a clear research question that can be tested using both methods.