Xavier Giró-i-Nieto

Konstantin Schürholt, Boris Knyazev, Xavier Giró-i-Nieto et al. · berkeley

Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and several sampling methods based on the topology of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform strong baselines as evaluated on several downstream tasks: initialization, ensemble sampling and transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.

Konstantin Schürholt, Diyar Taskiran, Boris Knyazev et al. · berkeley

In the last years, neural networks (NN) have evolved from laboratory environments to the state-of-the-art for many real-world problems. It was shown that NN models (i.e., their weights and biases) evolve on unique trajectories in weight space during training. Following, a population of such neural network models (referred to as model zoo) would form structures in weight space. We think that the geometry, curvature and smoothness of these structures contain information about the state of training and can reveal latent properties of individual models. With such model zoos, one could investigate novel approaches for (i) model analysis, (ii) discover unknown learning dynamics, (iii) learn rich representations of such populations, or (iv) exploit the model zoos for generative modelling of NN weights and biases. Unfortunately, the lack of standardized model zoos and available benchmarks significantly increases the friction for further research about populations of NNs. With this work, we publish a novel dataset of model zoos containing systematically generated and diverse populations of NN models for further research. In total the proposed model zoo dataset is based on eight image datasets, consists of 27 model zoos trained with varying hyperparameter combinations and includes 50'360 unique NN models as well as their sparsified twins, resulting in over 3'844'360 collected model states. Additionally, to the model zoo data we provide an in-depth analysis of the zoos and provide benchmarks for multiple downstream tasks. The dataset can be found at www.modelzoos.cc.

Laia Tarrés, Gerard I. Gállego, Amanda Duarte et al.

The advances in automatic sign language translation (SLT) to spoken languages have been mostly benchmarked with datasets of limited size and restricted domains. Our work advances the state of the art by providing the first baseline results on How2Sign, a large and broad dataset. We train a Transformer over I3D video features, using the reduced BLEU as a reference metric for validation, instead of the widely used BLEU score. We report a result of 8.03 on the BLEU score, and publish the first open-source implementation of its kind to promote further advances.

Konstantin Schürholt, Boris Knyazev, Xavier Giró-i-Nieto et al. · berkeley

Learning representations of neural network weights given a model zoo is an emerging and challenging area with many potential applications from model inspection, to neural architecture search or knowledge distillation. Recently, an autoencoder trained on a model zoo was able to learn a hyper-representation, which captures intrinsic and extrinsic properties of the models in the zoo. In this work, we extend hyper-representations for generative use to sample new model weights as pre-training. We propose layer-wise loss normalization which we demonstrate is key to generate high-performing models and a sampling method based on the empirical density of hyper-representations. The models generated using our methods are diverse, performant and capable to outperform conventional baselines for transfer learning. Our results indicate the potential of knowledge aggregation from model zoos to new models via hyper-representations thereby paving the avenue for novel research directions.

Pol Caselles, Eduard Ramon, Jaime Garcia et al.

Recovering the geometry of a human head from a single image, while factorizing the materials and illumination is a severely ill-posed problem that requires prior information to be solved. Methods based on 3D Morphable Models (3DMM), and their combination with differentiable renderers, have shown promising results. However, the expressiveness of 3DMMs is limited, and they typically yield over-smoothed and identity-agnostic 3D shapes limited to the face region. Highly accurate full head reconstructions have recently been obtained with neural fields that parameterize the geometry using multilayer perceptrons. The versatility of these representations has also proved effective for disentangling geometry, materials and lighting. However, these methods require several tens of input images. In this paper, we introduce SIRA, a method which, from a single image, reconstructs human head avatars with high fidelity geometry and factorized lights and surface materials. Our key ingredients are two data-driven statistical models based on neural fields that resolve the ambiguities of single-view 3D surface reconstruction and appearance factorization. Experiments show that SIRA obtains state of the art results in 3D head reconstruction while at the same time it successfully disentangles the global illumination, and the diffuse and specular albedos. Furthermore, our reconstructions are amenable to physically-based appearance editing and head model relighting.

Alvaro Budria, Laia Tarres, Gerard I. Gallego et al.

Significant progress has been made recently on challenging tasks in automatic sign language understanding, such as sign language recognition, translation and production. However, these works have focused on datasets with relatively few samples, short recordings and limited vocabulary and signing space. In this work, we introduce the novel task of sign language topic detection. We base our experiments on How2Sign, a large-scale video dataset spanning multiple semantic domains. We provide strong baselines for the task of topic detection and present a comparison between different visual features commonly used in the domain of sign language.

Laia Tarrés, Gerard I. Gàllego, Xavier Giró-i-Nieto et al.

This paper describes the system developed at the Universitat Politècnica de Catalunya for the Workshop on Machine Translation 2022 Sign Language Translation Task, in particular, for the sign-to-text direction. We use a Transformer model implemented with the Fairseq modeling toolkit. We have experimented with the vocabulary size, data augmentation techniques and pretraining the model with the PHOENIX-14T dataset. Our system obtains 0.50 BLEU score for the test set, improving the organizers' baseline by 0.38 BLEU. We remark the poor results for both the baseline and our system, and thus, the unreliability of our findings.

Jaume Ros, Margarita Geleta, Jordi Pons et al.

The field of steganography has experienced a surge of interest due to the recent advancements in AI-powered techniques, particularly in the context of multimodal setups that enable the concealment of signals within signals of a different nature. The primary objectives of all steganographic methods are to achieve perceptual transparency, robustness, and large embedding capacity - which often present conflicting goals that classical methods have struggled to reconcile. This paper extends and enhances an existing image-in-audio deep steganography method by focusing on improving its robustness. The proposed enhancements include modifications to the loss function, utilization of the Short-Time Fourier Transform (STFT), introduction of redundancy in the encoding process for error correction, and buffering of additional information in the pixel subconvolution operation. The results demonstrate that our approach outperforms the existing method in terms of robustness and perceptual transparency.

David Bonet, Daniel Mas Montserrat, Xavier Giró-i-Nieto et al.

Training deep learning models and performing hyperparameter tuning can be computationally demanding and time-consuming. Meanwhile, traditional machine learning methods like gradient-boosting algorithms remain the preferred choice for most tabular data applications, while neural network alternatives require extensive hyperparameter tuning or work only in toy datasets under limited settings. In this paper, we introduce HyperFast, a meta-trained hypernetwork designed for instant classification of tabular data in a single forward pass. HyperFast generates a task-specific neural network tailored to an unseen dataset that can be directly used for classification inference, removing the need for training a model. We report extensive experiments with OpenML and genomic data, comparing HyperFast to competing tabular data neural networks, traditional ML methods, AutoML systems, and boosting machines. HyperFast shows highly competitive results, while being significantly faster. Additionally, our approach demonstrates robust adaptability across a variety of classification tasks with little to no fine-tuning, positioning HyperFast as a strong solution for numerous applications and rapid model deployment. HyperFast introduces a promising paradigm for fast classification, with the potential to substantially decrease the computational burden of deep learning. Our code, which offers a scikit-learn-like interface, along with the trained HyperFast model, can be found at https://github.com/AI-sandbox/HyperFast.

Miriam Barrabes, Daniel Mas Montserrat, Margarita Geleta et al.

Data shift is a phenomenon present in many real-world applications, and while there are multiple methods attempting to detect shifts, the task of localizing and correcting the features originating such shifts has not been studied in depth. Feature shifts can occur in many datasets, including in multi-sensor data, where some sensors are malfunctioning, or in tabular and structured data, including biomedical, financial, and survey data, where faulty standardization and data processing pipelines can lead to erroneous features. In this work, we explore using the principles of adversarial learning, where the information from several discriminators trained to distinguish between two distributions is used to both detect the corrupted features and fix them in order to remove the distribution shift between datasets. We show that mainstream supervised classifiers, such as random forest or gradient boosting trees, combined with simple iterative heuristics, can localize and correct feature shifts, outperforming current statistical and neural network-based techniques. The code is available at https://github.com/AI-sandbox/DataFix.

Víctor Campos, Alexander Trott, Caiming Xiong et al.

Acquiring abilities in the absence of a task-oriented reward function is at the frontier of reinforcement learning research. This problem has been studied through the lens of empowerment, which draws a connection between option discovery and information theory. Information-theoretic skill discovery methods have garnered much interest from the community, but little research has been conducted in understanding their limitations. Through theoretical analysis and empirical evidence, we show that existing algorithms suffer from a common limitation -- they discover options that provide a poor coverage of the state space. In light of this, we propose 'Explore, Discover and Learn' (EDL), an alternative approach to information-theoretic skill discovery. Crucially, EDL optimizes the same information-theoretic objective derived from the empowerment literature, but addresses the optimization problem using different machinery. We perform an extensive evaluation of skill discovery methods on controlled environments and show that EDL offers significant advantages, such as overcoming the coverage problem, reducing the dependence of learned skills on the initial state, and allowing the user to define a prior over which behaviors should be learned. Code is publicly available at https://github.com/victorcampos7/edl.

Benet Oriol Sabat, Cristian Canton Ferrer, Xavier Giro-i-Nieto

This work addresses the challenge of hate speech detection in Internet memes, and attempts using visual information to automatically detect hate speech, unlike any previous work of our knowledge. Memes are pixel-based multimedia documents that contain photos or illustrations together with phrases which, when combined, usually adopt a funny meaning. However, hate memes are also used to spread hate through social networks, so their automatic detection would help reduce their harmful societal impact. Our results indicate that the model can learn to detect some of the memes, but that the task is far from being solved with this simple architecture. While previous work focuses on linguistic hate speech, our experiments indicate how the visual modality can be much more informative for hate speech detection than the linguistic one in memes. In our experiments, we built a dataset of 5,020 memes to train and evaluate a multi-layer perceptron over the visual and language representations, whether independently or fused. The source code and mode and models are available https://github.com/imatge-upc/hate-speech-detection .

Marc Górriz, Joseph Antony, Kevin McGuinness et al.

This work proposes a novel end-to-end convolutional neural network (CNN) architecture to automatically quantify the severity of knee osteoarthritis (OA) using X-Ray images, which incorporates trainable attention modules acting as unsupervised fine-grained detectors of the region of interest (ROI). The proposed attention modules can be applied at different levels and scales across any CNN pipeline helping the network to learn relevant attention patterns over the most informative parts of the image at different resolutions. We test the proposed attention mechanism on existing state-of-the-art CNN architectures as our base models, achieving promising results on the benchmark knee OA datasets from the osteoarthritis initiative (OAI) and multicenter osteoarthritis study (MOST). All code from our experiments will be publicly available on the github repository: https://github.com/marc-gorriz/KneeOA-CNNAttention

Panagiotis Linardos, Eva Mohedano, Juan Jose Nieto et al.

This paper investigates modifying an existing neural network architecture for static saliency prediction using two types of recurrences that integrate information from the temporal domain. The first modification is the addition of a ConvLSTM within the architecture, while the second is a conceptually simple exponential moving average of an internal convolutional state. We use weights pre-trained on the SALICON dataset and fine-tune our model on DHF1K. Our results show that both modifications achieve state-of-the-art results and produce similar saliency maps. Source code is available at https://git.io/fjPiB.

Marc Assens, Kevin McGuinness, Xavier Giro-i-Nieto et al.

We introduce SaltiNet, a deep neural network for scanpath prediction trained on 360-degree images. The model is based on a temporal-aware novel representation of saliency information named the saliency volume. The first part of the network consists of a model trained to generate saliency volumes, whose parameters are fit by back-propagation computed from a binary cross entropy (BCE) loss over downsampled versions of the saliency volumes. Sampling strategies over these volumes are used to generate scanpaths over the 360-degree images. Our experiments show the advantages of using saliency volumes, and how they can be used for related tasks. Our source code and trained models available at https://github.com/massens/saliency-360salient-2017.

Xavier Giro-i-Nieto, Nefeli Andreou, Anqi Liang et al.

Deep generative models have shown impressive progress in recent years, making it possible to produce high quality images with a simple text prompt or a reference image. However, state of the art technology does not yet meet the quality standards offered by traditional photographic methods. For this reason, production pipelines that use generated images often include a manual stage of image quality assessment (IQA). This process is slow and expensive, especially because of the low yield of automatically generated images that pass the quality bar. The IQA workload can be reduced by introducing an automatic pre-filtering stage, that will increase the overall quality of the images sent to review and, therefore, reduce the average cost required to obtain a high quality image. We present a formula that estimates the cost savings depending on the precision and pass yield of a generic IQA engine. This formula is applied in a use case of background inpainting, showcasing a significant cost saving of 51.61% obtained with a simple AutoML solution.

Daniel Mas Montserrat, David Bonet, Maria Perera et al.

The rapid expansion in the size of new datasets has created a need for fast and efficient parameter-learning techniques. Compressive learning is a framework that enables efficient processing by using random, non-linear features to project large-scale databases onto compact, information-preserving representations whose dimensionality is independent of the number of samples and can be easily stored, transferred, and processed. These database-level summaries are then used to decode parameters of interest from the underlying data distribution without requiring access to the original samples, offering an efficient and privacy-friendly learning framework. However, both the encoding and decoding techniques are typically randomized and data-independent, failing to exploit the underlying structure of the data. In this work, we propose a framework that meta-learns both the encoding and decoding stages of compressive learning methods by using neural networks that provide faster and more accurate systems than the current state-of-the-art approaches. To demonstrate the potential of the presented Compressive Meta-Learning framework, we explore multiple applications -- including neural network-based compressive PCA, compressive ridge regression, compressive k-means, and autoencoders.

Amanda Duarte, Samuel Albanie, Xavier Giró-i-Nieto et al.

Systems that can efficiently search collections of sign language videos have been highlighted as a useful application of sign language technology. However, the problem of searching videos beyond individual keywords has received limited attention in the literature. To address this gap, in this work we introduce the task of sign language retrieval with free-form textual queries: given a written query (e.g., a sentence) and a large collection of sign language videos, the objective is to find the signing video in the collection that best matches the written query. We propose to tackle this task by learning cross-modal embeddings on the recently introduced large-scale How2Sign dataset of American Sign Language (ASL). We identify that a key bottleneck in the performance of the system is the quality of the sign video embedding which suffers from a scarcity of labeled training data. We, therefore, propose SPOT-ALIGN, a framework for interleaving iterative rounds of sign spotting and feature alignment to expand the scope and scale of available training data. We validate the effectiveness of SPOT-ALIGN for learning a robust sign video embedding through improvements in both sign recognition and the proposed video retrieval task.

Eduard Ramon, Gil Triginer, Janna Escur et al.

Recent learning approaches that implicitly represent surface geometry using coordinate-based neural representations have shown impressive results in the problem of multi-view 3D reconstruction. The effectiveness of these techniques is, however, subject to the availability of a large number (several tens) of input views of the scene, and computationally demanding optimizations. In this paper, we tackle these limitations for the specific problem of few-shot full 3D head reconstruction, by endowing coordinate-based representations with a probabilistic shape prior that enables faster convergence and better generalization when using few input images (down to three). First, we learn a shape model of 3D heads from thousands of incomplete raw scans using implicit representations. At test time, we jointly overfit two coordinate-based neural networks to the scene, one modeling the geometry and another estimating the surface radiance, using implicit differentiable rendering. We devise a two-stage optimization strategy in which the learned prior is used to initialize and constrain the geometry during an initial optimization phase. Then, the prior is unfrozen and fine-tuned to the scene. By doing this, we achieve high-fidelity head reconstructions, including hair and shoulders, and with a high level of detail that consistently outperforms both state-of-the-art 3D Morphable Models methods in the few-shot scenario, and non-parametric methods when large sets of views are available.

Juan José Nieto, Roger Creus, Xavier Giro-i-Nieto

Pre-training Reinforcement Learning agents in a task-agnostic manner has shown promising results. However, previous works still struggle in learning and discovering meaningful skills in high-dimensional state-spaces, such as pixel-spaces. We approach the problem by leveraging unsupervised skill discovery and self-supervised learning of state representations. In our work, we learn a compact latent representation by making use of variational and contrastive techniques. We demonstrate that both enable RL agents to learn a set of basic navigation skills by maximizing an information theoretic objective. We assess our method in Minecraft 3D pixel maps with different complexities. Our results show that representations and conditioned policies learned from pixels are enough for toy examples, but do not scale to realistic and complex maps. To overcome these limitations, we explore alternative input observations such as the relative position of the agent along with the raw pixels.

Andreu Girbau, Xavier Giró-i-Nieto, Ignasi Rius et al.

Multiple object tracking faces several challenges that may be alleviated with trajectory information. Knowing the posterior locations of an object helps disambiguating and solving situations such as occlusions, re-identification, and identity switching. In this work, we show that trajectory estimation can become a key factor for tracking, and present TrajE, a trajectory estimator based on recurrent mixture density networks, as a generic module that can be added to existing object trackers. To provide several trajectory hypotheses, our method uses beam search. Also, relying on the same estimated trajectory, we propose to reconstruct a track after an occlusion occurs. We integrate TrajE into two state of the art tracking algorithms, CenterTrack [63] and Tracktor [3]. Their respective performances in the MOTChallenge 2017 test set are boosted 6.3 and 0.3 points in MOTA score, and 1.8 and 3.1 in IDF1, setting a new state of the art for the CenterTrack+TrajE configuration

Margarita Geleta, Cristina Punti, Kevin McGuinness et al.

Steganography comprises the mechanics of hiding data in a host media that may be publicly available. While previous works focused on unimodal setups (e.g., hiding images in images, or hiding audio in audio), PixInWav targets the multimodal case of hiding images in audio. To this end, we propose a novel residual architecture operating on top of short-time discrete cosine transform (STDCT) audio spectrograms. Among our results, we find that the residual audio steganography setup we propose allows independent encoding of the hidden image from the host audio without compromising quality. Accordingly, while previous works require both host and hidden signals to hide a signal, PixInWav can encode images offline -- which can be later hidden, in a residual fashion, into any audio signal. Finally, we test our scheme in a lab setting to transmit images over airwaves from a loudspeaker to a microphone verifying our theoretical insights and obtaining promising results.

Ioannis Kazakos, Carles Ventura, Miriam Bellver et al.

Recent advances in deep learning have brought significant progress in visual grounding tasks such as language-guided video object segmentation. However, collecting large datasets for these tasks is expensive in terms of annotation time, which represents a bottleneck. To this end, we propose a novel method, namely SynthRef, for generating synthetic referring expressions for target objects in an image (or video frame), and we also present and disseminate the first large-scale dataset with synthetic referring expressions for video object segmentation. Our experiments demonstrate that by training with our synthetic referring expressions one can improve the ability of a model to generalize across different datasets, without any additional annotation cost. Moreover, our formulation allows its application to any object detection or segmentation dataset.

Oscar Mañas, Alexandre Lacoste, Xavier Giro-i-Nieto et al.

Remote sensing and automatic earth monitoring are key to solve global-scale challenges such as disaster prevention, land use monitoring, or tackling climate change. Although there exist vast amounts of remote sensing data, most of it remains unlabeled and thus inaccessible for supervised learning algorithms. Transfer learning approaches can reduce the data requirements of deep learning algorithms. However, most of these methods are pre-trained on ImageNet and their generalization to remote sensing imagery is not guaranteed due to the domain gap. In this work, we propose Seasonal Contrast (SeCo), an effective pipeline to leverage unlabeled data for in-domain pre-training of remote sensing representations. The SeCo pipeline is composed of two parts. First, a principled procedure to gather large-scale, unlabeled and uncurated remote sensing datasets containing images from multiple Earth locations at different timestamps. Second, a self-supervised algorithm that takes advantage of time and position invariance to learn transferable representations for remote sensing applications. We empirically show that models trained with SeCo achieve better performance than their ImageNet pre-trained counterparts and state-of-the-art self-supervised learning methods on multiple downstream tasks. The datasets and models in SeCo will be made public to facilitate transfer learning and enable rapid progress in remote sensing applications.

Lucas Ventura, Amanda Duarte, Xavier Giro-i-Nieto

Recent work have addressed the generation of human poses represented by 2D/3D coordinates of human joints for sign language. We use the state of the art in Deep Learning for motion transfer and evaluate them on How2Sign, an American Sign Language dataset, to generate videos of signers performing sign language given a 2D pose skeleton. We evaluate the generated videos quantitatively and qualitatively showing that the current models are not enough to generated adequate videos for Sign Language due to lack of detail in hands.

Miriam Bellver, Carles Ventura, Carina Silberer et al.

The task of video object segmentation with referring expressions (language-guided VOS) is to, given a linguistic phrase and a video, generate binary masks for the object to which the phrase refers. Our work argues that existing benchmarks used for this task are mainly composed of trivial cases, in which referents can be identified with simple phrases. Our analysis relies on a new categorization of the phrases in the DAVIS-2017 and Actor-Action datasets into trivial and non-trivial REs, with the non-trivial REs annotated with seven RE semantic categories. We leverage this data to analyze the results of RefVOS, a novel neural network that obtains competitive results for the task of language-guided image segmentation and state of the art results for language-guided VOS. Our study indicates that the major challenges for the task are related to understanding motion and static actions.

Miriam Bellver, Amaia Salvador, Jordi Torres et al.

Image segmentation methods are usually trained with pixel-level annotations, which require significant human effort to collect. The most common solution to address this constraint is to implement weakly-supervised pipelines trained with lower forms of supervision, such as bounding boxes or scribbles. Another option are semi-supervised methods, which leverage a large amount of unlabeled data and a limited number of strongly-labeled samples. In this second setup, samples to be strongly-annotated can be selected randomly or with an active learning mechanism that chooses the ones that will maximize the model performance. In this work, we propose a sample selection approach to decide which samples to annotate for semi-supervised instance segmentation. Our method consists in first predicting pseudo-masks for the unlabeled pool of samples, together with a score predicting the quality of the mask. This score is an estimate of the Intersection Over Union (IoU) of the segment with the ground truth mask. We study which samples are better to annotate given the quality score, and show how our approach outperforms a random selection, leading to improved performance for semi-supervised instance segmentation with low annotation budgets.

Amanda Duarte, Shruti Palaskar, Lucas Ventura et al.

One of the factors that have hindered progress in the areas of sign language recognition, translation, and production is the absence of large annotated datasets. Towards this end, we introduce How2Sign, a multimodal and multiview continuous American Sign Language (ASL) dataset, consisting of a parallel corpus of more than 80 hours of sign language videos and a set of corresponding modalities including speech, English transcripts, and depth. A three-hour subset was further recorded in the Panoptic studio enabling detailed 3D pose estimation. To evaluate the potential of How2Sign for real-world impact, we conduct a study with ASL signers and show that synthesized videos using our dataset can indeed be understood. The study further gives insights on challenges that computer vision should address in order to make progress in this field. Dataset website: http://how2sign.github.io/

Maria Gonzalez-i-Calabuig, Carles Ventura, Xavier Giró-i-Nieto

Video object segmentation can be understood as a sequence-to-sequence task that can benefit from the curriculum learning strategies for better and faster training of deep neural networks. This work explores different schedule sampling and frame skipping variations to significantly improve the performance of a recurrent architecture. Our results on the car class of the KITTI-MOTS challenge indicate that, surprisingly, an inverse schedule sampling is a better option than a classic forward one. Also, that a progressive skipping of frames during training is beneficial, but only when training with the ground truth masks instead of the predicted ones. Source code and trained models are available at http://imatge-upc.github.io/rvos-mots/.

Benet Oriol, Jordi Luque, Ferran Diego et al.

In this work, we propose an effective approach for training unique embedding representations by combining three simultaneous modalities: image and spoken and textual narratives. The proposed methodology departs from a baseline system that spawns a embedding space trained with only spoken narratives and image cues. Our experiments on the EPIC-Kitchen and Places Audio Caption datasets show that introducing the human-generated textual transcriptions of the spoken narratives helps to the training procedure yielding to get better embedding representations. The triad speech, image and words allows for a better estimate of the point embedding and show an improving of the performance within tasks like image and speech retrieval, even when text third modality, text, is not present in the task.

Alba Herrera-Palacio, Carles Ventura, Carina Silberer et al.

The goal of this work is to segment the objects in an image that are referred to by a sequence of linguistic descriptions (referring expressions). We propose a deep neural network with recurrent layers that output a sequence of binary masks, one for each referring expression provided by the user. The recurrent layers in the architecture allow the model to condition each predicted mask on the previous ones, from a spatial perspective within the same image. Our multimodal approach uses off-the-shelf architectures to encode both the image and the referring expressions. The visual branch provides a tensor of pixel embeddings that are concatenated with the phrase embeddings produced by a language encoder. Our experiments on the RefCOCO dataset for still images indicate how the proposed architecture successfully exploits the sequences of referring expressions to solve a pixel-wise task of instance segmentation.

Mariona Caros, Maite Garolera, Petia Radeva et al.

With people living longer than ever, the number of cases with dementia such as Alzheimer's disease increases steadily. It affects more than 46 million people worldwide, and it is estimated that in 2050 more than 100 million will be affected. While there are not effective treatments for these terminal diseases, therapies such as reminiscence, that stimulate memories from the past are recommended. Currently, reminiscence therapy takes place in care homes and is guided by a therapist or a carer. In this work, we present an AI-based solution to automatize the reminiscence therapy, which consists in a dialogue system that uses photos as input to generate questions. We run a usability case study with patients diagnosed of mild cognitive impairment that shows they found the system very entertaining and challenging. Overall, this paper presents how reminiscence therapy can be automatized by using machine learning, and deployed to smartphones and laptops, making the therapy more accessible to every person affected by dementia.

Eduard Ramon, Guillermo Ruiz, Thomas Batard et al.

This work proposes novel hyperparameter-free losses for single view 3D reconstruction with morphable models (3DMM). We dispense with the hyperparameters used in other works by exploiting geometry, so that the shape of the object and the camera pose are jointly optimized in a sole term expression. This simplification reduces the optimization time and its complexity. Moreover, we propose a novel implicit regularization technique based on random virtual projections that does not require additional 2D or 3D annotations. Our experiments suggest that minimizing a shape reprojection error together with the proposed implicit regularization is especially suitable for applications that require precise alignment between geometry and image spaces, such as augmented reality. We evaluate our losses on a large scale dataset with 3D ground truth and publish our implementations to facilitate reproducibility and public benchmarking in this field.

Miriam Bellver, Amaia Salvador, Jordi Torres et al.

Methods that move towards less supervised scenarios are key for image segmentation, as dense labels demand significant human intervention. Generally, the annotation burden is mitigated by labeling datasets with weaker forms of supervision, e.g. image-level labels or bounding boxes. Another option are semi-supervised settings, that commonly leverage a few strong annotations and a huge number of unlabeled/weakly-labeled data. In this paper, we revisit semi-supervised segmentation schemes and narrow down significantly the annotation budget (in terms of total labeling time of the training set) compared to previous approaches. With a very simple pipeline, we demonstrate that at low annotation budgets, semi-supervised methods outperform by a wide margin weakly-supervised ones for both semantic and instance segmentation. Our approach also outperforms previous semi-supervised works at a much reduced labeling cost. We present results for the Pascal VOC benchmark and unify weakly and semi-supervised approaches by considering the total annotation budget, thus allowing a fairer comparison between methods.

Amanda Duarte, Francisco Roldan, Miquel Tubau et al.

Speech is a rich biometric signal that contains information about the identity, gender and emotional state of the speaker. In this work, we explore its potential to generate face images of a speaker by conditioning a Generative Adversarial Network (GAN) with raw speech input. We propose a deep neural network that is trained from scratch in an end-to-end fashion, generating a face directly from the raw speech waveform without any additional identity information (e.g reference image or one-hot encoding). Our model is trained in a self-supervised approach by exploiting the audio and visual signals naturally aligned in videos. With the purpose of training from video data, we present a novel dataset collected for this work, with high-quality videos of youtubers with notable expressiveness in both the speech and visual signals.

Carles Ventura, Miriam Bellver, Andreu Girbau et al.

Multiple object video object segmentation is a challenging task, specially for the zero-shot case, when no object mask is given at the initial frame and the model has to find the objects to be segmented along the sequence. In our work, we propose a Recurrent network for multiple object Video Object Segmentation (RVOS) that is fully end-to-end trainable. Our model incorporates recurrence on two different domains: (i) the spatial, which allows to discover the different object instances within a frame, and (ii) the temporal, which allows to keep the coherence of the segmented objects along time. We train RVOS for zero-shot video object segmentation and are the first ones to report quantitative results for DAVIS-2017 and YouTube-VOS benchmarks. Further, we adapt RVOS for one-shot video object segmentation by using the masks obtained in previous time steps as inputs to be processed by the recurrent module. Our model reaches comparable results to state-of-the-art techniques in YouTube-VOS benchmark and outperforms all previous video object segmentation methods not using online learning in the DAVIS-2017 benchmark. Moreover, our model achieves faster inference runtimes than previous methods, reaching 44ms/frame on a P100 GPU.

Patrick Bilic, Patrick Christ, Hongwei Bran Li et al.

In this work, we report the set-up and results of the Liver Tumor Segmentation Benchmark (LiTS), which was organized in conjunction with the IEEE International Symposium on Biomedical Imaging (ISBI) 2017 and the International Conferences on Medical Image Computing and Computer-Assisted Intervention (MICCAI) 2017 and 2018. The image dataset is diverse and contains primary and secondary tumors with varied sizes and appearances with various lesion-to-background levels (hyper-/hypo-dense), created in collaboration with seven hospitals and research institutions. Seventy-five submitted liver and liver tumor segmentation algorithms were trained on a set of 131 computed tomography (CT) volumes and were tested on 70 unseen test images acquired from different patients. We found that not a single algorithm performed best for both liver and liver tumors in the three events. The best liver segmentation algorithm achieved a Dice score of 0.963, whereas, for tumor segmentation, the best algorithms achieved Dices scores of 0.674 (ISBI 2017), 0.702 (MICCAI 2017), and 0.739 (MICCAI 2018). Retrospectively, we performed additional analysis on liver tumor detection and revealed that not all top-performing segmentation algorithms worked well for tumor detection. The best liver tumor detection method achieved a lesion-wise recall of 0.458 (ISBI 2017), 0.515 (MICCAI 2017), and 0.554 (MICCAI 2018), indicating the need for further research. LiTS remains an active benchmark and resource for research, e.g., contributing the liver-related segmentation tasks in \url{http://medicaldecathlon.com/}. In addition, both data and online evaluation are accessible via \url{www.lits-challenge.com}.

Amaia Salvador, Michal Drozdzal, Xavier Giro-i-Nieto et al.

People enjoy food photography because they appreciate food. Behind each meal there is a story described in a complex recipe and, unfortunately, by simply looking at a food image we do not have access to its preparation process. Therefore, in this paper we introduce an inverse cooking system that recreates cooking recipes given food images. Our system predicts ingredients as sets by means of a novel architecture, modeling their dependencies without imposing any order, and then generates cooking instructions by attending to both image and its inferred ingredients simultaneously. We extensively evaluate the whole system on the large-scale Recipe1M dataset and show that (1) we improve performance w.r.t. previous baselines for ingredient prediction; (2) we are able to obtain high quality recipes by leveraging both image and ingredients; (3) our system is able to produce more compelling recipes than retrieval-based approaches according to human judgment. We make code and models publicly available.

Víctor Campos, Xavier Giro-i-Nieto, Jordi Torres

Evolution Strategies (ES) emerged as a scalable alternative to popular Reinforcement Learning (RL) techniques, providing an almost perfect speedup when distributed across hundreds of CPU cores thanks to a reduced communication overhead. Despite providing large improvements in wall-clock time, ES is data inefficient when compared to competing RL methods. One of the main causes of such inefficiency is the collection of large batches of experience, which are discarded after each policy update. In this work, we study how to perform more than one update per batch of experience by means of Importance Sampling while preserving the scalability of the original method. The proposed method, Importance Weighted Evolution Strategies (IW-ES), shows promising results and is a first step towards designing efficient ES algorithms.

Marc Assens, Xavier Giro-i-Nieto, Kevin McGuinness et al.

We introduce PathGAN, a deep neural network for visual scanpath prediction trained on adversarial examples. A visual scanpath is defined as the sequence of fixation points over an image defined by a human observer with its gaze. PathGAN is composed of two parts, the generator and the discriminator. Both parts extract features from images using off-the-shelf networks, and train recurrent layers to generate or discriminate scanpaths accordingly. In scanpath prediction, the stochastic nature of the data makes it very difficult to generate realistic predictions using supervised learning strategies, but we adopt adversarial training as a suitable alternative. Our experiments prove how PathGAN improves the state of the art of visual scanpath prediction on the iSUN and Salient360! datasets. Source code and models are available at https://imatge-upc.github.io/pathgan/

Panagiotis Linardos, Eva Mohedano, Monica Cherto et al.

This work adapts a deep neural model for image saliency prediction to the temporal domain of egocentric video. We compute the saliency map for each video frame, firstly with an off-the-shelf model trained from static images, secondly by adding a a convolutional or conv-LSTM layers trained with a dataset for video saliency prediction. We study each configuration on EgoMon, a new dataset made of seven egocentric videos recorded by three subjects in both free-viewing and task-driven set ups. Our results indicate that the temporal adaptation is beneficial when the viewer is not moving and observing the scene from a narrow field of view. Encouraged by this observation, we compute and publish the saliency maps for the EPIC Kitchens dataset, in which viewers are cooking. Source code and models available at https://imatge-upc.github.io/saliency-2018-videosalgan/

Daniel Fojo, Víctor Campos, Xavier Giro-i-Nieto

Adaptive Computation Time for Recurrent Neural Networks (ACT) is one of the most promising architectures for variable computation. ACT adapts to the input sequence by being able to look at each sample more than once, and learn how many times it should do it. In this paper, we compare ACT to Repeat-RNN, a novel architecture based on repeating each sample a fixed number of times. We found surprising results, where Repeat-RNN performs as good as ACT in the selected tasks. Source code in TensorFlow and PyTorch is publicly available at https://imatge-upc.github.io/danifojo-2018-repeatrnn/

Zheng Shou, Junting Pan, Jonathan Chan et al.

We aim to tackle a novel task in action detection - Online Detection of Action Start (ODAS) in untrimmed, streaming videos. The goal of ODAS is to detect the start of an action instance, with high categorization accuracy and low detection latency. ODAS is important in many applications such as early alert generation to allow timely security or emergency response. We propose three novel methods to specifically address the challenges in training ODAS models: (1) hard negative samples generation based on Generative Adversarial Network (GAN) to distinguish ambiguous background, (2) explicitly modeling the temporal consistency between data around action start and data succeeding action start, and (3) adaptive sampling strategy to handle the scarcity of training data. We conduct extensive experiments using THUMOS'14 and ActivityNet. We show that our proposed methods lead to significant performance gains and improve the state-of-the-art methods. An ablation study confirms the effectiveness of each proposed method.

Didac Surís, Amanda Duarte, Amaia Salvador et al.

The increasing amount of online videos brings several opportunities for training self-supervised neural networks. The creation of large scale datasets of videos such as the YouTube-8M allows us to deal with this large amount of data in manageable way. In this work, we find new ways of exploiting this dataset by taking advantage of the multi-modal information it provides. By means of a neural network, we are able to create links between audio and visual documents, by projecting them into a common region of the feature space, obtaining joint audio-visual embeddings. These links are used to retrieve audio samples that fit well to a given silent video, and also to retrieve images that match a given a query audio. The results in terms of Recall@K obtained over a subset of YouTube-8M videos show the potential of this unsupervised approach for cross-modal feature learning. We train embeddings for both scales and assess their quality in a retrieval problem, formulated as using the feature extracted from one modality to retrieve the most similar videos based on the features computed in the other modality.

Amaia Salvador, Miriam Bellver, Victor Campos et al.

We present a recurrent model for semantic instance segmentation that sequentially generates binary masks and their associated class probabilities for every object in an image. Our proposed system is trainable end-to-end from an input image to a sequence of labeled masks and, compared to methods relying on object proposals, does not require post-processing steps on its output. We study the suitability of our recurrent model on three different instance segmentation benchmarks, namely Pascal VOC 2012, CVPPP Plant Leaf Segmentation and Cityscapes. Further, we analyze the object sorting patterns generated by our model and observe that it learns to follow a consistent pattern, which correlates with the activations learned in the encoder part of our network. Source code and models are available at https://imatge-upc.github.io/rsis/

Miriam Bellver, Kevis-Kokitsi Maninis, Jordi Pont-Tuset et al.

A fully automatic technique for segmenting the liver and localizing its unhealthy tissues is a convenient tool in order to diagnose hepatic diseases and assess the response to the according treatments. In this work we propose a method to segment the liver and its lesions from Computed Tomography (CT) scans using Convolutional Neural Networks (CNNs), that have proven good results in a variety of computer vision tasks, including medical imaging. The network that segments the lesions consists of a cascaded architecture, which first focuses on the region of the liver in order to segment the lesions on it. Moreover, we train a detector to localize the lesions, and mask the results of the segmentation network with the positive detections. The segmentation architecture is based on DRIU, a Fully Convolutional Network (FCN) with side outputs that work on feature maps of different resolutions, to finally benefit from the multi-scale information learned by different stages of the network. The main contribution of this work is the use of a detector to localize the lesions, which we show to be beneficial to remove false positives triggered by the segmentation network. Source code and models are available at https://imatge-upc.github.io/liverseg-2017-nipsws/ .

Eva Mohedano, Kevin McGuinness, Xavier Giro-i-Nieto et al.

This work explores attention models to weight the contribution of local convolutional representations for the instance search task. We present a retrieval framework based on bags of local convolutional features (BLCF) that benefits from saliency weighting to build an efficient image representation. The use of human visual attention models (saliency) allows significant improvements in retrieval performance without the need to conduct region analysis or spatial verification, and without requiring any feature fine tuning. We investigate the impact of different saliency models, finding that higher performance on saliency benchmarks does not necessarily equate to improved performance when used in instance search tasks. The proposed approach outperforms the state-of-the-art on the challenging INSTRE benchmark by a large margin, and provides similar performance on the Oxford and Paris benchmarks compared to more complex methods that use off-the-shelf representations. The source code used in this project is available at https://imatge-upc.github.io/salbow/

Victor Campos, Brendan Jou, Xavier Giro-i-Nieto et al.

Recurrent Neural Networks (RNNs) continue to show outstanding performance in sequence modeling tasks. However, training RNNs on long sequences often face challenges like slow inference, vanishing gradients and difficulty in capturing long term dependencies. In backpropagation through time settings, these issues are tightly coupled with the large, sequential computational graph resulting from unfolding the RNN in time. We introduce the Skip RNN model which extends existing RNN models by learning to skip state updates and shortens the effective size of the computational graph. This model can also be encouraged to perform fewer state updates through a budget constraint. We evaluate the proposed model on various tasks and show how it can reduce the number of required RNN updates while preserving, and sometimes even improving, the performance of the baseline RNN models. Source code is publicly available at https://imatge-upc.github.io/skiprnn-2017-telecombcn/ .

Delia Fernandez, Alejandro Woodward, Victor Campos et al.

The increasing availability of affect-rich multimedia resources has bolstered interest in understanding sentiment and emotions in and from visual content. Adjective-noun pairs (ANP) are a popular mid-level semantic construct for capturing affect via visually detectable concepts such as "cute dog" or "beautiful landscape". Current state-of-the-art methods approach ANP prediction by considering each of these compound concepts as individual tokens, ignoring the underlying relationships in ANPs. This work aims at disentangling the contributions of the `adjectives' and `nouns' in the visual prediction of ANPs. Two specialised classifiers, one trained for detecting adjectives and another for nouns, are fused to predict 553 different ANPs. The resulting ANP prediction model is more interpretable as it allows us to study contributions of the adjective and noun components. Source code and models are available at https://imatge-upc.github.io/affective-2017-musa2/ .

Xunyu Lin, Victor Campos, Xavier Giro-i-Nieto et al.

This paper introduces an unsupervised framework to extract semantically rich features for video representation. Inspired by how the human visual system groups objects based on motion cues, we propose a deep convolutional neural network that disentangles motion, foreground and background information. The proposed architecture consists of a 3D convolutional feature encoder for blocks of 16 frames, which is trained for reconstruction tasks over the first and last frames of the sequence. A preliminary supervised experiment was conducted to verify the feasibility of proposed method by training the model with a fraction of videos from the UCF-101 dataset taking as ground truth the bounding boxes around the activity regions. Qualitative results indicate that the network can successfully segment foreground and background in videos as well as update the foreground appearance based on disentangled motion features. The benefits of these learned features are shown in a discriminative classification task, where initializing the network with the proposed pretraining method outperforms both random initialization and autoencoder pretraining. Our model and source code are publicly available at https://imatge-upc.github.io/unsupervised-2017-cvprw/ .