Abstract: Hand sign language recognition from video is a challenging research area in computer vision, which performance is affected by hand occlusion, fast hand movement, illumination changes, or background complexity, just to mention a few. In recent years, deep learning approaches have achieved state-of-the-art results in the field, though previous challenges are not completely solved. In this work, we propose a novel deep learning-based pipeline architecture for efficient automatic hand sign language recognition using Single Shot Detector (SSD), 2D Convolutional Neural Network (2DCNN), 3D Convolutional Neural Network (3DCNN), and Long Short-Term Memory (LSTM) from RGB input videos. We use a CNN-based model which estimates the 3D hand keypoints from 2D input frames. After that, we connect these estimated keypoints to build the hand skeleton by using midpoint algorithm. In order to obtain a more discriminative representation of hands, we project 3D hand skeleton into three views surface images. We further employ the heatmap image of detected keypoints as input for refinement in a stacked fashion. We apply 3DCNNs on the stacked features of hand, including pixel level, multi-view hand skeleton, and heatmap features, to extract discriminant local spatio-temporal features from these stacked inputs. The outputs of the 3DCNNs are fused and fed to a LSTM to model long-term dynamics of hand sign gestures. Analyzing 2DCNN vs. 3DCNN using different number of stacked inputs into the network, we demonstrate that 3DCNN better capture spatio-temporal dynamics of hands. To the best of our knowledge, this is the first time that this multi-modal and multi-view set of hand skeleton features are applied for hand sign language recognition. Furthermore, we present a new large-scale hand sign language dataset, namely RKS-PERSIANSIGN, including 10′000 RGB videos of 100 Persian sign words. Evaluation results of the proposed model on three datasets, NYU, First-Person, and RKS-PERSIANSIGN, indicate that our model outperforms state-of-the-art models in hand sign language recognition, hand pose estimation, and hand action recognition.

Abstract: Isolated hand sign language recognition from video is a challenging research area in computer vision. Some of the most important challenges in this area include dealing with hand occlusion, fast hand movement, illumination changes, or background complexity. While most of the state-of-the-art results in the field have been achieved using deep learning-based models, the previous challenges are not completely solved. In this paper, we propose a hand pose aware model for isolated hand sign language recognition using deep learning approaches from two input modalities, RGB and depth videos. Four spatial feature types: pixel-level, flow, deep hand, and hand pose features, fused from both visual modalities, are input to LSTM for temporal sign recognition. While we use Optical Flow (OF) for flow information in RGB video inputs, Scene Flow (SF) is used for depth video inputs. By including hand pose features, we show a consistent performance improvement of the sign language recognition model. To the best of our knowledge, this is the first time that this discriminant spatiotemporal features, benefiting from the hand pose estimation features and multi-modal inputs, are fused for isolated hand sign language recognition. We perform a step-by-step analysis of the impact in terms of recognition performance of the hand pose features, different combinations of the spatial features, and different recurrent models, especially LSTM and GRU. Results on four public datasets confirm that the proposed model outperforms the current state-of-the-art models on Montalbano II, MSR Daily Activity 3D, and CAD-60 datasets with a relative accuracy improvement of 1.64%, 6.5%, and 7.6%. Furthermore, our model obtains a competitive results on isoGD dataset with only 0.22% margin lower than the current state-of-the-art model.

Abstract: In this work we target the problem of hate speech detection in multimodal publications formed by a text and an image. We gather and annotate a large scale dataset from Twitter, MMHS150K, and propose different models that jointly analyze textual and visual information for hate speech detection, comparing them with unimodal detection. We provide quantitative and qualitative results and analyze the challenges of the proposed task. We find that, even though images are useful for the hate speech detection task, current multimodal models cannot outperform models analyzing only text. We discuss why and open the field and the dataset for further research.

Abstract: Machine learning experimentation under controlled scenarios and standard datasets is necessary to compare algorithms performance by evaluating all of them in the same setup. However, experimentation on how those algorithms perform on unconstrained data and applied tasks to solve real world problems is also a must to ascertain how that research can contribute to our society.
In this dissertation we experiment with the latest computer vision and natural language processing algorithms applying them to multimodal scene interpretation. Particularly, we research on how image and text understanding can be jointly exploited to address real world problems, focusing on learning from Social Media data.
We address several tasks that involve image and textual information, discuss their characteristics and offer our experimentation conclusions. First, we work on detection of scene text in images. Then, we work with Social Media posts, exploiting the captions associated to images as supervision to learn visual features, which we apply to multimodal semantic image retrieval. Subsequently, we work with geolocated Social Media images with associated tags, experimenting on how to use the tags as supervision, on location sensitive image retrieval and on exploiting location information for image tagging. Finally, we work on a specific classification problem of Social Media publications consisting on an image and a text: Multimodal hate speech classification.

Abstract: Salient object segmentation in videos is generally broken up in a video segmentation part and a saliency assignment part. Recently, object proposals, which are used to segment the image, have had significant impact on many computer vision applications, including image segmentation, object detection, and recently saliency detection in still images. However, their usage has not yet been evaluated for salient object segmentation in videos. Therefore, in this paper, we investigate the application of object proposals to salient object segmentation in videos. In addition, we propose a new motion feature derived from the optical flow structure tensor for video saliency detection. Experiments on two standard benchmark datasets for video saliency show that the proposed motion feature improves saliency estimation results, and that object proposals are an efficient method for salient object segmentation. Results on the challenging SegTrack v2 and Fukuchi benchmark data sets show that we significantly outperform the state-of-the-art.

Abstract: This paper proposes a novel CycleGAN architecture for thermal image super-resolution, together with a large dataset consisting of thermal images at different resolutions. The dataset has been acquired using three thermal cameras at different resolutions, which acquire images from the same scenario at the same time. The thermal cameras are mounted in rig trying to minimize the baseline distance to make easier the registration problem.
The proposed architecture is based on ResNet6 as a Generator and PatchGAN as Discriminator. The novelty on the proposed unsupervised super-resolution training (CycleGAN) is possible due to the existence of aforementioned thermal images—images of the same scenario with different resolutions. The proposed approach is evaluated in the dataset and compared with classical bicubic interpolation. The dataset and the network are available.

Abstract: Social networks have attracted the attention of psychologists, as the behavior of users can be used to assess personality traits, and to detect sentiments and critical mental situations such as depression or suicidal tendencies. Recently, the increasing amount of image uploads to social networks has shifted the focus from text to image-based personality assessment. However, obtaining the ground-truth requires giving personality questionnaires to the users, making the process very costly and slow, and hindering research on large populations. In this paper, we demonstrate that it is possible to predict which images are most associated with each personality trait of the OCEAN personality model, without requiring ground-truth personality labels. Namely, we present a weakly supervised framework which shows that the personality scores obtained using specific images textually associated with particular personality traits are highly correlated with scores obtained using standard text-based personality questionnaires. We trained an OCEAN trait model based on Convolutional Neural Networks (CNNs), learned from 120K pictures posted with specific textual hashtags, to infer whether the personality scores from the images uploaded by users are consistent with those scores obtained from text. In order to validate our claims, we performed a personality test on a heterogeneous group of 280 human subjects, showing that our model successfully predicts which kind of image will match a person with a given level of a trait. Looking at the results, we obtained evidence that personality is not only correlated with text, but with image content too. Interestingly, different visual patterns emerged from those images most liked by persons with a particular personality trait: for instance, pictures most associated with high conscientiousness usually contained healthy food, while low conscientiousness pictures contained injuries, guns, and alcohol. These findings could pave the way to complement text-based personality questionnaires with image-based questions.

Abstract: During the last years, graph-based representations are experiencing a growing usage in visual recognition and retrieval due to their ability to capture both structural and appearance-based information. Thus, they provide a greater representational power than classical statistical frameworks. However, graph-based representations leads to high computational complexities usually dealt by graph embeddings or approximated matching techniques. Despite their representational power, they are very sensitive to noise and small variations of the input image. With the aim to cope with the time complexity and the variability present in the generated graphs, in this paper we propose to construct a novel hierarchical graph representation. Graph clustering techniques adapted from social media analysis have been used in order to contract a graph at different abstraction levels while keeping information about the topology. Abstract nodes attributes summarise information about the contracted graph partition. For the proposed representations, a coarse-to-fine matching technique is defined. Hence, small graphs are used as a filtering before more accurate matching methods are applied. This approach has been validated in real scenarios such as classification of colour images or retrieval of handwritten words (i.e. word spotting).